Picking-Point Localization Algorithm for Citrus Fruits Based on Improved YOLOv8 Model

Abstract

The citrus picking-point localization is critical for automatic citrus harvesting. Due to the complex citrus growing environment and the limitations of devices, the efficient citrus picking-point localization method becomes a hot research topic. This study designs a novel and efficient workflow for citrus picking-point localization, named as CPPL. The CPPL is achieved based on two stages, namely the detection stage and the segmentation stage. For the detection stage, we define the KD-YOLOP to accurately detect citrus fruits to quickly localize the initial picking region. The KD-YOLOP is defined based on a knowledge distillation learning and a model pruning to reduce the computational cost while having a competitive accuracy. For the segmentation stage, we define the RG-YOLO-seg to efficiently segment the citrus branches to compute the picking points. The RG-YOLO-seg is proposed by introducing the RGNet to extract efficient features and using the GSNeck to fuse multi-scale features. Therefore, by using knowledge distillation, model pruning, and a lightweight model for branch segmentation, the proposed CPPL achieves accurate real-time localization of citrus picking points. We conduct extensive experiments to evaluate our method; many results show that the proposed CPPL outperforms the current methods and achieves adequate accuracy. It provides an efficient and robust novel method for real-time citrus harvesting in practical agricultural applications.

1. Introduction

Fruit harvesting is a crucial procedure in the fruit industry. With the rapid growing of global citrus industry, labor issues are becoming increasingly acute for most of countries that produce citrus [1,2]. For example, the citrus industry in Belize faces a critical labor shortage due to high costs and stringent requirements for migrant workers, leading to significant reductions in citrus harvested and potential economic losses of up to $15 million [3]. In the U.S., the department of agriculture is attempting to introduce a program to solve the labor shortage of citrus industry by providing incentives [4]. In China, citrus production grew by 7% and reached over 64 million metric tons, and will continually increase [5]. Therefore, there is an urgent need to increase productivity for the citrus industry. In most regions, citrus harvesting is still predominantly manual. Due to its seasonality, high yields, and short harvest window, manual labor struggles to keep up [6]. Automatic harvesting is one of the keys to modernizing agriculture; automatic fruit picking has become as a prominent research focus, demonstrating the potential to address labor shortages [7].

In the research of automatic fruit-picking robots, the design of machine vision algorithms based on deep learning has become a hot topic in recent years. The key problem in automatic fruit picking is to accurately and stably localize the picking points of fruits based on its end effector and the target fruit [8,9]. Many factors may affect the accuracy of picking point localization, such as complex growing environments of fruits, fruit clustering and occlusion, illumination conditions and weathers [10]. Despite these challenges, automatic fruit picking must meet requirements like precise and fast positioning, fast picking, adaptability to different environments, less destruction, and so on [11]. To address these challenges, researchers have focused on developing more efficient and practical methods for picking point localization. In addition to the influences from robot machines, fruits, and their harvesting environments, the present picking point localization methods primarily rely on object detection and instance segmentation algorithms. Based on object detection, instance segmentation, and post processing, there are generally three types of picking point localization methods, including detection-based methods, segmentation-based methods, and detection–segmentation-based methods.

The detection-based and segmentation-based methods usually optimize the fruit detection algorithm or fruit instance segmentation algorithm first, and then introduce some prior assumptions based on the growth of fruits to localize picking points. In one instance with detection-based methods, Zhang et al. [12] combined a laser active-vision system with the YOLOv5 model to locate apples on trees with an average precision mAP of 92.86%. By obtaining accurate two-dimensional pixel coordinates and converting them to three-dimensional world coordinates, they defined fruit coordinates as picking points. Similar approaches have been applied to various fruits, including citrus [13,14], apples [12,15,16,17], pineapples [11], grapes [18,19], tomatoes [20,21], litchi [22], nectarines [23], kiwis [24], passion fruits [25], and so on. For the segmentation-based methods, Li et al. [26] developed an optimized one-stage instance segmentation model for litchi harvesting and combined the segmentation result with statistical information for litchi clusters’ picking point localization to reach an accuracy of 82% for locating litchi. Similar strategies have been applied to blueberries [27,28], grapes [29,30], waxberries [31], strawberries [32,33], tomatoes [34] and so on. These methods demonstrate the versatility of detection-based and segmentation-based approaches for various fruit types and harvesting scenarios. However, these methods are limited by more complex environments in real-world orchards, resulting in poor accuracy for picking point localization due to their reliance on detection or segmentation results and their own post-processing methods.

To further improve the efficiency and accuracy to localize the picking points of various fruits, researchers develop many detection–segmentation-based methods that integrated an object detection algorithm and instance segmentation algorithm at the same time. For instance, Huang et al. [35] introduced an optimized network for grape cluster localization, combining detection and instance segmentation within a Mask R-CNN framework. Their results were 1.4% and 1.5% higher than the original model. Similarly, Yan et al. [36] proposed an improved algorithm based on the YOLOv8 model and achieved an mAP of 98.3% on apple recognition and an mAP of 81.6% on branch and trunk segmentation, enabling simultaneous apple fruit detection and branch segmentation for precise picking point localization. These detection–segmentation-based approaches take advantage of the detection and segmentation, providing detailed spatial and structural information about fruits and their surroundings, thereby improving the localization process. Therefore, many other methods are proposed to be applied to fruits such as grapes [37], apples [38], kiwifruit [39], citrus [40], strawberry [41], and so on. Although all of these methods promote the localization accuracy, however, they need detection and segmentation results at the same time, so multiple neural networks need to be owned. These studies are overly focused on improving individual detection or segmentation metrics, without optimizing the design in conjunction with the hardware operation. This is a critical factor for the low computational power of picking robots.

To address these issues, we tend to design the entire citrus-picking process in collaboration with the hardware partner and optimize models mainly in terms of latency. To improve the efficiency of automatic citrus picking, we study the real picking process of worker in a real orchard. As shown in the Figure 1, the intuitive picking process of a worker includes: moving close to the tree, selecting a cluster of citrus fruits, and cutting off the branch connected to the citrus fruits. In their actual work, workers do not always cut the branches connected to the fruit; sometimes they also need to cut the branches that obstruct the fruit. Therefore, we select the shear cutter as our end effector and leave the potential for routine pruning during the growing period. We select the IntelSense Depth Camera for vision. For the baseline model, we chose the YOLOv8, because it has been the most widely used in the last two years of the research field in automatic fruit picking, with the aim of making our results more reproducible and popularizing the merits of our approach.

Unlike most of previous works, this study does not completely concentrate on improving the evaluation metrics of the detection or segmentation model. However, to optimize the latency performance, there are three methods used in this study, including optimized modules, knowledge distillation, and model pruning. For the optimized modules, we mainly select the lightweight modules GSConv [42] and RepGhost [43], although it may lower the accuracy performance. The knowledge distillation and model pruning are two commonly used real-time optimization methods in the field of deep learning. The knowledge distillation is a kind of efficient transfer learning method used for model compression and optimization [44]. The basic idea is to acquire knowledge from a large model (called the teacher model) by training a small model (called the student model). This process aims to reduce the size of the model while maintaining its performance as much as possible, making it suitable for resource-constrained environments such as mobile devices or embedded systems. The model pruning is to used for cutting off redundant neurons, it can effectively reduce the number of parameters of the model without affecting the performance too much. In deep neural networks, different neurons have different effects on the performance. For a freshly trained neural network, there may be a large number of redundant parameters, where a large number of neurons have activation values that tend to be close to 0, meaning that these neurons do not participate in the training or inference. In conclusion, all of our optimization methods are first and foremost designed to be deployed on real devices, then to balance the performance of accuracy and latency.

To lower the latency of the entire system process, we proposed a novel workflow for citrus picking-point localization (CPPL). As shown in Figure 2 and Algorithm 1, we design the CPPL based on considerations of optimizing the integration of hardware and software. The CPPL is constructed by defining the KD-YOLOP to detect citrus fruits, the spatial clustering algorithm to extract RoI, the RG-YOLO-seg to segment the citrus branches, and the efficient self-designed image algorithm to localize picking points. First, the KD-YOLOP is employed to obtain coordinates of citrus fruits. Second, the spatial clustering algorithm is employed on the detection results of citrus fruits. Based on the clusters of citrus fruits, multiple RoI are extracted. Third, the RG-YOLO-seg is employed to obtain the segmentation result of citrus branches. With the detection result and segmentation pixels, we obtain coordinates from each RoI. Many experiments demonstrated that the CPPL produced real-time localization of pixel-level picking points for citrus fruits, providing a significant improvement over existing methods. Our main contributions are as follows:

• (1). We propose an efficient and practical workflow for citrus picking-point localization (CPPL). The CPPL satisfies the low computational requirements of automatic picking robot and adapts to real orchard environments. The CPPL consists of our proposed detection (KD-YOLOP) and instance segmentation (RG-YOLO-seg) algorithms, achieving accurate real-time localization of citrus picking-point localization.

• (2). We develop the KD-YOLOP, a novel citrus fruit detection model that balances accuracy and latency. The model employs a knowledge distillation framework, using YOLOv8l as the teacher model and YOLOv8n as the student model, with channel-wise loss ensuring effective knowledge alignment. Through four rounds of pruning, the KD-YOLOP significantly reduces parameters compared to the existing detection methods while maintaining competitive accuracy.

• (3). We develop the RG-YOLO-seg, a lightweight instance segmentation model optimized for citrus branch segmentation. This model incorporates the RGNet as the backbone and the GSNeck as the neck. The RGNet utilizes the efficient RepGhost module for lightweight feature extraction, while the GSNeck employs GSConv for efficient multi-scale feature fusion. The proposed RG-YOLO-seg reduces computational complexity while maintaining strong segmentation performance.

• (4). We conduct extensive experiments to evaluate our proposed methods, dividing them into quantitative and qualitative analysis. The results demonstrate that our proposed algorithms outperform others, while the proposed CPPL achieves adequate performance.

Algorithm 1: System processing of CPPL.

[Figure omitted. See PDF] [Figure omitted. See PDF]

2. Materials and Methods

2.1. Dataset

The images used in this study were captured in citrus orchards located in Nanning, Guangxi Province, China. All data were collected during the daytime, including ripe and unripe fruits, as shown in Figure 3. To simulate the view of camera in the robot, we collected images with a camera-to-fruit distance ranging from 0.5 m to 1.0 m, under sunny, cloudy, and windy circumstances. Due to the complexity of the orchard environment and in consideration of personal safety, we did not conduct data collection under extreme conditions (e.g., rainy days or nighttime). All images were preprocessed to a resolution of 1280 × 720 pixels. The annotation tool is Labelme (https://github.com/wkentaro/labelme (accessed on 9 October 2022)). Two datasets were constructed: the Citrus Fruits Detection Dataset (CFDD) and Citrus Branches Segmentation Dataset (CBSD). As shown in Figure 4, we employed data augmentation, including flipping, translation, brightness adjustment, saturation adjustment, adding noise and rotation, to expand the original dataset. As shown in the Figure 5, we plot the position of citrus fruits in scatter and heat map, after the data augmentation, the distribution of spatial locations has become more balanced. The CFDD dataset includes 9000 labeled images, and the CBSD dataset includes 28,000 labeled images. As presented in Table 1, we divided the CFDD and the CBSD datasets into training and validation subsets.

2.2. Workflow for Citrus Picking-Point Localization

The proposed novel and efficient workflow for citrus picking-point localization (CPPL) is shown in Figure 2. The figure also shows our hardware, where the end effector of the robotic arm is a self-designed shear cutter, and the camera mounted on the robotic arm is the Intel RealSense Depth Camera D435i. In the first stage, the captured image is transmitted to our proposed algorithm KD-YOLOP for citrus fruit detection. Then, the detection results are used to obtain the region of interest by a spatial cluster algorithm. In the second stage, the extracted regions of interest are transmitted to our proposed RG-YOLO-seg for citrus branch instance segmentation. Finally, the detection results of citrus fruits and the segmentation results of citrus branches are used to localize picking points of all extracted regions of interest through our designed image algorithm and logic. The coordinates of picking points are transferred into three dimensions by accessing the depth of the captured image; subsequently, the picking robot acts on these coordinates one by one. To illustrate this more clearly, we present the following pseudo-code in Algorithm 1.

In particular, the picking-point localization step can be optimized for parallel computing; the loop traversal presented above is for ease of understanding.

2.3. KD-YOLOP for Citrus Fruit Detection Based on Knowledge Distillation and Model Pruning

Figure 6 shows the constructing of the proposed KD-YOLOP for citrus detection; it is divided in two steps: knowledge distillation (KD) and model pruning (MP). For the KD step, YOLOv8l model is trained on the CFDD dataset to serve as the teacher network, while the YOLOv8n model is designated as the student network. Subsequently, the student network is trained for knowledge distillation using the CFDD dataset and the output of the teacher network. For the MD step, the student model in the last step is pruned an arbitrary number of times. The more iterations of pruning, the fewer parameters the KD-YOLOP retains.

• (1). Knowledge Distillation

Knowledge distillation (KD) is one of the model compression and acceleration techniques used in the development of efficient deep models [44]. It is typically employed to transfer most of the performance of a large-scale neural network to a smaller one. Besides the original target detection loss function, an additional knowledge distillation loss function is introduced to measure the difference between the teacher network and the student network. By minimizing the knowledge distillation loss function, the student network can better learn the “knowledge” of the teacher network, aligning their feature parameters. In this paper, Channel-Wise Distillation (CWD) (as shown in Equation (1)) is chosen as the distillation loss function, which is achieved by normalizing the activation maps of each channel of the two networks and transforming them into probabilistic distributions, and then using the Kullback–Leibler (KL) divergence (as shown in Equation (2)) to measure the difference between the two models, and finally reducing the difference between the two models through the back-propagation of the models. The CWD loss and KL divergence function are shown in Equations (1) and (2):

(1)$\varphi (\mathsf{\Phi }\left(y^T\right),\mathsf{\Phi }\left(y^S\right))=\varphi (\mathsf{\Phi }\left(y_c^T\right),\mathsf{\Phi }\left(y_c^S\right)),$

(2)$\mathsf{\Phi }\left(y_c\right)={\displaystyle \frac{exp\left(\frac{y_{c,i}}{\tau }\right)}{{\sum }_{i=1}^{W}exp\left(\frac{y_{c,i}}{\tau }\right)}}.$

In Equation (1), T and S denote the teacher network and the student network, respectively. $y^T$ and $y^S$ denote the activation of these two networks, respectively. $\mathsf{\Phi }(·)$ is the transformation function, which is responsible for converting the network activation values into a probabilistic distribution. In Equation (2), $c=1,2,3\dots$ denotes the channel, i denotes the index of each channel, the $\tau$ is the temperature hyper-parameter, and $y_{c,i}$ denotes the outputs of each channel in the student network. $\varphi (·)$ is the KL scatter function, which is used to measure the difference between the channel distributions of the two networks.

• (2). Model Pruning

The lower part of Figure 6 illustrates the pruning framework for the distilled model (the student network from the previous step). To effectively reduce the computational cost while maintaining the accuracy, model pruning is employed; this technique aims to eliminate redundant neurons that do not contribute to generating meaningful results. Model pruning mainly includes sparse training, pruning, and fine-tuning. First, the sparse training reduces computational cost and memory by zeroing out less important weights. Then, the pruning removes unimportant weights. Finally, the fine-tuning adjusts the remaining parameters to improve model performance. In this section, we employ four iterations of pruning on the distilled student model. As illustrated in Euqation (3), we introduce L1 regularization terms to each Batch Normalization (BN) layer of the distilled student model to facilitate sparse training.

(3)$L=\sum _{x,y}l(f(x,W),y)+\lambda \sum _{\gamma \subseteq \tau }g\left(\gamma \right),$

In Equation (3), l represents the original loss function of the model, where $(x,y)$ corresponds to the training data and labels. The W denotes the model’s trainable parameters, while $g(·)$ represents the penalization term applied to the scaling factor. The parameter $\lambda$ serves as the balance factor, controlling the trade-off between the original loss and the regularization term. Lastly, $\gamma$ is the scaling factor introduced during the sparse training.

In the pruning step, we focus on the Conv2d and BatchNorm2d layers within the Conv and C2f modules, as well as the Linear layer in the decoupling header. The fine-tuning step ensures the model relearns any lost features or adapts to the newly pruned structure. Based on these procedures, we perform four iterations of model pruning. Finally, we select the model weights that strike the best balance between the computational cost and accuracy, and designate this model as KD-YOLOP.

2.4. The RG-YOLO-seg for Citrus Branch Segmentation Based on Lightweight Modules

Using the proposed KD-YOLOP described in Section 2.3, citrus fruits are detected in the captured image. Subsequently, regions of interest are extracted based on the detection results (the red bounding boxes of citrus fruits in Figure 7). As illustrated in Figure 7, citrus picking-point localization is achieved using the RG-YOLO-seg, which performs instance segmentation of the citrus branches.

The RG-YOLO-seg framework comprises three main components: the backbone (RGNet), the neck (GSNeck), and the segment head (Decode Head). The RGNet, designed using the RepGhost module, serves as the backbone and enhances feature extraction while maintaining a lightweight structure. The GSNeck, adopting principles from GSConv, is integrated as the neck to facilitate efficient multi-scale feature processing. The Decode Head, adapted from the YOLOv8-seg, is included to perform instance segmentation, effectively complementing the proposed the RGNet and the GSNeck.

As shown in Figure 7, the RGNet consists of five layers. Except for the first and fifth layers, each layer includes a Conv Module (ConvModule) and an RG module (RGModule). The ConvModule integrates a 2D convolution, a BatchNorm layer, and a SiLU activation function. The RGModule employs the RepGhost module, a reparameterization-based design that decomposes standard convolution into primary and cheap operations to reduce the computational cost of feature extraction. By incorporating two RepGhost modules connected via jump-linking, the RGModule promotes feature reuse. This structure enables the generation of diverse feature maps with reduced computational cost, thereby expanding network capacity and improving overall feature extraction efficiency. Additionally, the SPPF module, included as a part of RGNet, serves as an advanced version of the Spatial Pyramid Polling (SPP) module, converting feature maps of varying sizes into fixed-length feature vectors to further optimize representation efficiency.

The GSNeck is integrated into the RG-YOLO-seg to balance semantic information preservation and model efficiency. This module, as illustrated in the green box of Figure 7, employs the GSConv operation exclusively in the neck, where feature map dimensions are already compressed. The strategic placement reduces the model’s parameter count while retaining representational depth. By preserving essential channel connections without introducing additional transformations, the GSNeck effectively enhances the network’s overall efficiency and real-time performance, achieving an optimal balance between innovation and computational cost.

3. Experiments

This section outlines the experimental setup and analysis. First, the experimental environment, the parameter configurations, and and the evaluation metric are introduced in Section 3.1. Second, the performance evaluation and the compared experiments are conducted to validate the proposed methods in Section 3.2 and Section 3.3. Third, the qualitative analysis are conducted in Section 3.4. Finally, the ablation studies are performed to analyze the contribution of key components in Section 3.5.

3.1. Experimental Environment and Parameter Settings

The experiments in this study are conducted using an NVIDIA GeForce RTX 3090 GPU paired with Intel Xeon Silver 4214R processor. The operating system is Ubuntu 18.04 and its GPU driver is CUDA 11.4, with 128 GB of RAM, and 24 GB of GPU memory. The software is developed using Python 3.7.13 and PyTorch 1.7.1.

During the training process, the batch size is set to 48, the number of training epochs is 100, and the learning rate is fixed at 0.001. The dropout rate of KD-YOLOP and RG-YOLO-seg is set to 0. The optimization strategy of KD-YOLOP and RG-YOLO-seg is the Stochastic Gradient Descent (SGD). For the loss function, this study employs three types to facilitate model training and pruning: Binary Cross-Entropy (BCE) Loss for classification branch, Distribution Focal Loss for regression branch, and Complete Intersection over Union (CIoU) Loss for the regression branch. During back-propagation, these loss functions are weighted with a ratio of 7.5:0.5:1.5. Additionally, Channel-Wise Distillation (CWD) Loss is utilized to minimize differences between the teacher and student networks during knowledge distillation training.

The evaluation metric for citrus fruit detection and citrus branch segmentation is mean average precision (mAP) based on the known COCO dataset [45]. In the detection experiments, we use the mAP_50:95. In the segmentation experiments, we set Intersection over Union (IoU) thresholds to 0.35 (mAP₃5). The CFDD and the CBSD datasets are used for citrus fruit detection and citrus branch instance segmentation experimental training, respectively. During the training process, the image pairs were cropped and enhanced, with the final input image size standardized to 640 × 640 pixels.

3.2. Quantitative Analysis and Evaluation of KD-YOLOP

The experiments for KD-YOLOP include in two steps: the knowledge distillation and the model pruning. These two experiments are based on the CFDD dataset.

Table 2 shows the results of the knowledge distillation (KD) step experiments of KD-YOLOP. The RT-DETR [46], YOLOv5s, YOLOv8l, YOLOv8n, YOLOv10s [47] and YOLOv11s [48] are selected as comparison methods, as they are the most widely used detection models in applications. Although the detection accuracy of YOLOv8l is 8.7% higher than that of YOLOv8n. However, the parameter size and computational cost of YOLOv8l are 14 times and 20 times that of YOLOv8n, respectively. The KD-YOLOP maintains the same number of parameters and computational cost as the YOLOv8n, but the detection accuracy is improved by 6.6%. In addition, we compare our method with YOLOv5s, YOLOv10s, and YOLOv11s, which are the most commonly used and newest detection model in industry, and the RT-DETR, the latest real-time Transformer detector. The detection accuracy of KD-YOLOP is 1.8% higher than that of YOLOv5s model, outperforming it in terms of the latency performance at the same time. Compared with YOLOv10s and YOLOv11s, the KD-YOLOP is 2.6% and 3.4% lower, repsectively. However, the parameter size and computational cost of them is approximately 3 times of KD-YOLOP. Compared with RT-DETR, it is 1.6% higher. The model parameter size of KD-YOLOP is 3.0 M, and the computational cost is 8.1 GFLOPs, both the model scale and computational cost of KD-YOLOP are smaller.

After the knowledge distillation, we conduct the model pruning experiments. The more iterations of pruning, the fewer accuracy and parameters the KD-YOLOP retains. To achieve best balance between accuracy and latency, we conduct four iterations. Compared with the distilled model in the last step, the model parameter size is reduced by 56.7%, the computational cost is reduced by 48.1%, and the accuracy is reduced by 10.2%.

3.3. Quantitative Analysis and Evaluation of the RG-YOLO-seg

The proposed RG-YOLO-seg algorithm was evaluated on the CBSD dataset. As shown in Table 3, we compare the RG-YOLO-seg with four other lightweight instance segmentation models (YOLOv8n-seg, YOLOv8l-seg, Yolact [49], Mask-RCNN [50] and SparseInst [51]) in terms of average precision, parameter size (M), and GFLOPs. Compared with YOLOv8n-seg, which utilizes the lightweight MobileNetV3 [52], our method demonstrates a similar number of parameters and computational cost. However, it outperforms the YOLOv8n-seg by achieving 6.3% higher segmentation accuracy. These results highlight the competitiveness of our method in balancing accuracy and latency.

3.4. Qualitative Analysis

We conduct a qualitative analysis of the KD-YOLOP and the RG-YOLO-seg based on visualized results of the CFDD dataset and the CBSD dataset. In addition, we conduct an error analysis on images outside the two datasets. We divide our experiments in two circumstances, unripe fruits and ripe fruits; the first is used to simulate our algorithm applied in a fruit thinning scenario and the second is used to simulate our algorithm in a harvesting scenario.

• (1). Citrus fruit detection on the CFDD dataset of unripe fruits

Figure 8 presents experimental comparisons on the CFDD dataset of unripe fruits. The first, second, and third columns show the results of YOLOv8n, YOLOv8l, and KD-YOLOP, respectively. The second row is a magnification of the area specified in the first row of images, and the fourth row is a magnification of the area specified in the third row. The second row is a magnification of the area specified in the first row of images, and the fourth row is a magnification of the area specified in the third row. As shown in the second and the fourth rows, in the areas of fruit-to-fruit overlap and fruit-to-leaf occlusion, both the distilled KD-YOLOP model and the teacher model exhibit a more accurate localization of citrus fruits. This suggests that the KD-YOLOP is more adaptive in handling complex details, further validating the successful transfer of knowledge from the teacher network to the student network during the distillation process.

• (2). Citrus fruit detection on the CFDD dataset of ripe fruits

Figure 9 presents experimental comparisons on the CFDD dataset of ripe fruits. The first, second, and third columns show the results of YOLOv8n, YOLOv8l and KD-YOLOP, respectively. The second row is a magnification of the area specified in the first row of images, and the fourth row is a magnification of the area specified in the third row. Comparing the second and the forth rows in Figure 9, the proposed KD-YOLOP successfully detects more citrus fruits than others, particularly in the circumstance of the fruit–leaf occlusion.

• (3). Citrus branch instance segmentation on the CBSD dataset of unripe fruits

Figure 10 presents experimental comparisons on the CBSD dataset of unripe fruits. The first, second, and third columns show the results of YOLOv8n, YOLOv8l, and KD-YOLOP, respectively. As shown in the first, second, and forth rows in Figure 10, the proposed RG-YOLO-seg successfully segments the most complete results of citrus branches. This means that the RG-YOLO-seg is more suitable for the proposed CPPL, as it may combines with the cluster boxes of citrus, further promoting the accuracy of the CPPL.

• (4). Citrus branch instance segmentation on the CBSD dataset of ripe fruits

Figure 11 presents experimental comparisons on the CBSD dataset of unripe fruits. The first, second, and third columns show the results of YOLOv8n, YOLOv8l, and KD-YOLOP, respectively. For the ripe fruits, the color difference between citrus fruits and citrus branches is shown. Except for the incorrect prediction made by Mask-RCNN, the other two methods successfully segment the branch connected to the citrus fruit. Comparing the YOLOv8n-seg and the RG-YOLO-seg, as shown in the forth row in Figure 11, the proposed RG-YOLO-seg successfully segment branches in the circumstance of leaf–branch occlusion.

• (5). Error Analysis

As shown in Figure 12 and Figure 13, we select several error situations while testing our models on unlabeled images. Figure 12 shows the detection result of KD-YOLOP. Observing the detection results of the ripe fruits shown in the first row, we can see that for the fruits that are far away, there is an occlusion, which produces an undetectable area. For the second row, we can see that the model incorrectly identifies leaves as unripe fruits. This phenomenon may be due to the annotation habit; for the ripe fruits, we ignore the fruits with too large a distance, and for the unripe fruits, we carefully label all the fruits, even if they are almost completely occluded. To solve these errors, the most effective way is to revise our dataset. Figure 13 shows the segmentation result of RG-YOLO-seg. For ripe and unripe fruits, in both cases, the model fails to recognize all the pixels of the branches. When the model sees the trunk of a tree at a distance, the trunk is incorrectly recognized as a branch. This problem can be solved intuitively by combining the depth camera. When it comes to the scenario of unripe fruit, the model fails to recognize green branches (as illustrated in the third scene of Figure 13). To solve this kind of similarity interference, one potential direction for improvement is to enhance the extraction of high-frequency features.

3.5. Ablation Study

To evaluate the effectiveness of our proposed methods, we conduct many ablation studies on the KD-YOLOP and the RG-YOLO-seg.

• (1). KD-YOLOP for Citrus Fruit Detection

In this paper, ablation experiments on knowledge distillation and model pruning for KD-YOLOP were conducted using the CFDD dataset to verify the effectiveness of both methods. As shown in Table 4, we defined YOLOv8n as the baseline model and YOLOv8l as the teacher network during the knowledge distillation learning. KD-YOLOP (1) was selected as the representative model for comparison.

The average precision (AP) of the baseline model is 85.7%, with a parameter count of 3.0 M and a computational cost of 8.1 GFLOPs. When knowledge distillation is applied, the model’s AP increases by 6.6%, while the number of parameters remains unchanged. In contrast, when only the pruning module is applied, the model experiences a 6.3% decrease in AP after four pruning iterations, with reductions of 56.7% in parameters and 49.3% in computational cost. However, when both distillation and pruning techniques are combined in our proposed KD-YOLOP, the model achieves an AP of 85.3% while maintaining its lightweight nature. This highlights the effectiveness of the proposed methods.

• (2). The RG-YOLO-seg for Citrus Branch Instance Segmentation

We conduct ablation experiments of the RGNet and the GSNeck modules on the CBSD dataset to verify the effectiveness of the two modules, respectively. We use YOLOv8n-seg as the baseline, with an AP of 77.6%, a parameter count of 3.3 M, and a computational cost of 12.0 GFLOPs. When the RGNet is used as a lightweight backbone, the AP decrease by 1.4% but the parameters and the computational volume decreased by 6% and 5%, respectively. When we use GSConv to construct the lightweight neck, the AP of the model decreased by 0.9%, but the number of parameters and computational cost decreased by 9% and 7.5%, respectively. In contrast, when both the RGNet and the GSNeck are used to construct the RG-YOLO-seg model, the AP of the model decreased by 0.4%, while the number of parameters and computational cost decrease by 12% and 11.7%. Although each lightweight improvement causes a certain loss of AP, the latency performance of the model was effectively improved, and, as shown in Figure 10 and Figure 11, the mild decrease in segmentation AP did not cause any obvious influence (Table 5).

3.6. Picking-Point Localization Results and Analysis

As shown in Figure 14, we test our proposed CPPL on various scenarios, the red dots inside the yellow circles indicate the picking-point locations. These scenarios can be categorized by the relative position of a targeted citrus fruit and its connected branch, including the middle scene (as shown in the first column in Figure 14), the scene to the left (as shown in the second column in Figure 14), the scene to the right (as shown in the first and the second rows of the third column in Figure 14), and the special down–up scene (as shown in the second and the third columns of the third row in Figure 14). These results confirm that the proposed method maintains excellent localization accuracy while ensuring model efficiency.

4. Discussion

Our optimization of the model focuses on lightweighting, with the goal of achieving the best balance between accuracy and latency. In our practice, it shows that a few percentile changes in detection or segmentation metrics do not affect the identification of picking points that much. In particular, the tests in the real orchard are completely different from those in the laboratory where the values are compared. When the detection or segmentation metrics reach a certain high value, the bottleneck of citrus picking-point localization turns to the latency of these detection and segmentation models, and our designed algorithmic logic uses the results of citrus fruit detection and citrus branch segmentation. The design logic of the entire system, however, is continuously modified and iterated upon during field testing. Therefore, the most meaningful feature of our optimized deep learning model is that it provides a sufficiently stable accuracy at the lowest possible latency (fewer parameters and computation cost) when used as part of the CPPL.

In the field of automatic fruit picking, various methodologies have been explored, including classical computer vision techniques, 3D reconstruction-based methods, and LiDAR-based methods. Each of these approaches has its unique strengths and applications, particularly in leveraging diverse sensor technologies. The classical computer vision techniques have been widely used due to their ability to process 2D image data effectively before the boost of deep learning methods, offering solutions for object detection and segmentation. However, their reliance on monocular visual information may limit their capability in more complex spatial perception tasks. On the other hand, 3D reconstruction-based methods excel in spatial understanding and depth perception, enabling the accurate modeling of complex environments. Similarly, LiDAR-based methods, with their precise distance measurement capabilities, are particularly advantageous for tasks requiring high accuracy in environmental mapping and obstacle detection.

We acknowledge the thoughtful utilization of these diverse sensors in the aforementioned techniques and their considerable contributions to agricultural automation. Despite their advantages, it is important to note that computer vision (CV)-based approaches have seen remarkable advancements in recent years, driven by the rapid development of deep learning algorithms and improved computational efficiency. These advancements have made CV-based methods increasingly robust and scalable for real-world applications.

At this stage, our work focuses on CV-based solutions due to their stability and adaptability in various scenarios. Nevertheless, we recognize the potential of integrating 3D reconstruction-based and LiDAR-based techniques to further enhance the capabilities of automatic fruit-picking systems. Future research will aim to explore such integrations, leveraging the complementary strengths of these methods to advance the field.

5. Conclusions

In this study, the proposed workflow for citrus picking-point localization (CPPL) is to address the limitations of automatic citrus harvesting in a real-world orchard. We collaborated with the hardware partner, and the the main idea of our strategy was to balance the accuracy and latency of algorithms used in the automatic harvest processing. Therefore, we propose the KD-YOLOP for citrus fruit detection and the RG-YOLO-seg for citrus branch instance segmentation. Using the knowledge distillation and model pruning for the KD-YOLOP, and the lightweight model for the RG-YOLO-seg, we finally achieved real-time citrus picking-point localization. We constructed the Citrus Fruits Detection Dataset (CFDD) and the Citrus Branches Segmentation Dataset (CBSD) for our extensive experiments. We conducted extensive experiments based on the CFDD and the CBSD datasets to evaluate our method; many results show that the proposed CPPL outperforms the mentioned methods and achieves adequate accuracy. It provides an efficient and robust novel method for real-time citrus harvesting in agricultural applications.

To the best of our knowledge and searched literature, the YOLOv8 model is widely used in all kinds of applications. To be more generalizable, the proposed KD-YOLOP and the RG-YOLO-seg are based on YOLOv8. As the proposed CPPL is scalable and flexible in adapting different models, the performance of the CPPL can be easily improved by more advanced models in the future. This feature also enhances the generalization of our method to the automatic harvesting of other fruits.

Author Contributions

Conceptualization, W.J. and Y.L. (Yun Liang); methodology, W.J.; software, W.J. and Y.L. (Yunfan Liu); validation, W.J. and Y.L. (Yun Liang); investigation, W.J. and Z.W.; resources, Y.L. (Yun Liang); data curation, W.J., R.Z. and Z.W.; writing—original draft preparation, W.J.; writing—review and editing, Y.L. (Yun Liang); supervision, Y.L. (Yun Liang); project administration, W.J. and Y.L. (Yun Liang); funding acquisition, Y.L. (Yun Liang). All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

Not applicable.

Data Availability Statement

The raw data supporting the conclusions of this article will be made available by the authors on request.

Acknowledgments

We sincerely thank Guangdong Tecsun Science & Technology Co., Ltd. for providing hardware support, and Guangxi Minggu Agriculture Co., Ltd. and Guangxi Nanning Wanjin Agriculture Co., Ltd. for providing the experimental orchard.

Conflicts of Interest

The authors declare no conflicts of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. Citrus-picking process.

Figure 2. Workflow for citrus picking-point localization.

Figure 3. Raw image data of citrus.

Figure 4. Examples of data augmentation.

Figure 5. Changes in statistical characteristics of data augmentation.

Figure 6. Training process of KD-YOLOP and citrus-cluster generator.

Figure 7. An overview of the proposed RG-YOLO-seg.

Figure 8. Comparisons on the CFDD dataset of unripe fruits.

Figure 9. Comparisons on the CFDD dataset of ripe fruits.

Figure 10. Comparisons on the CBSD dataset of unripe fruits.

Figure 11. Comparisons on the CBSD dataset of ripe fruits.

Figure 12. Error analysis of citrus fruit detection.

Figure 13. Error analysis of citrus branch segmentation.

Figure 14. Visualized results of citrus picking-point localization.

Table 1

Citrus Fruits Detection Dataset (CFDD) and Citrus Branches Segmentation Dataset (CBSD).

Dataset	Class	Train	Validate
CFDD	unripe	3500	1000
	ripe	3500	100
	unripe + ripe	7000	2000
CBSD	unripe	10,500	3500
	ripe	10,500	3500
	unripe + ripe	21,000	7000

Table 2

Comparisons of knowledge distillation experiments for citrus detection models.

Model	Learning Rate	mAP50:95 (%)	Parameter (M)	GFLOPs
RT-DETR	0.0001	90.7	19.7	60.1
YOLOv5s	0.001	90.5	7.2	15.8
YOLOv8l	0.001	94.4	43.6	164.0
YOLOv8n	0.001	85.7	3.0	8.1
YOLOv10s	0.001	94.9	8.0	24.8
YOLOv11s	0.001	95.7	9.0	21.5
KD-YOLOP (ours)	0.001	92.3	3.0	8.1

Table 3

Comparisons of different segmentation models.

Model	Learning Rate	Backbone	AP35 (%)	Parameter (M)	GFLOPs
YOLOv8l-seg	0.001	CSPDarkNet	81.1	45.9	220.1
YOLOv8n-seg	0.001	MobileNetv3	70.8	2.6	9.5
Mask-RCNN	0.01	ResNet50	70.3	101.5	491.4
Yolact	5 × 10−5	ResNet50	64.2	34.7	186.2
SparseInst	5 × 10−6	ResNet50	76.3	33.9	181.7
RG-YOLO-seg (ours)	0.001	RGNet	77.1	2.9	10.6

Table 4

The KD-YOLOP ablation experiments.

Model	Learning Rate	Epochs	mAP50:95 (%)	Parameter (M)	GFLOPs
YOLOv8n	0.001	100	85.7	3.0	8.1
YOLOv8n + KD	0.001	100	92.3	3.0	8.1
YOLOv8n + Prune	0.001	100	79.4	1.3	4.1
KD-YOLOP	0.001	100	85.3	1.9	5.6

Table 5

The RG-YOLO-seg ablation experiments.

Model	Learning Rate	Epochs	AP35 (%)	Parameter	GFLOPs
YOLOv8n-seg	0.001	100	77.6	3.3	12.0
YOLOv8n-seg + RGNet	0.001	100	77.0	3.1	11.4
YOLOv8n-seg + GSNeck	0.001	100	77.5	3.0	11.1
RG-YOLO-seg	0.001	100	77.1	2.9	10.6

References

1. Qi, L.; Qi, C. Status Quo and Development Trend of World’s Citrus Industry. Agric. Outlook; 2016; 12, pp. 46-52. [DOI: https://dx.doi.org/10.3969/j.issn.1673-3908.2016.12.010]

2. FreshPlaza. Global Market Report; FreshPlaza: Citrus, TX, USA, 2025.

3. Humes, A. Severe Labour Shortage to Harvest Crop Plagues the Citrus Industry; Up to $15 Million in Losses Possible. 2022; Available online: https://www.breakingbelizenews.com/2022/03/04/severe-labour-shortage-to-harvest-crop-plagues-the-citrus-industry-up-to-15-million-in-losses-possible/ (accessed on 1 October 2024).

4. Citrus Industry Magazine. Program Provides Incentives to Address Labor Challenges. 2023; Available online: https://citrusindustry.net/2023/10/13/program-provides-incentives-to-address-labor-challenges/ (accessed on 1 October 2024).

5. FreshFruitPortal. Chinese Citrus Production to Increase Slightly, Despite Challenges. 2025; Available online: https://www.freshfruitportal.com/news/2025/01/07/chinese-citrus-production-to-increase-slightly-despite-challenges/ (accessed on 1 October 2024).

6. Hou, G.; Chen, H.; Jiang, M.; Niu, R. An Overview of the Application of Machine Vision in Recognition and Localization of Fruit and Vegetable Harvesting Robots. Agriculture; 2023; 13, 1814. [DOI: https://dx.doi.org/10.3390/agriculture13091814]

7. Xin, Q.; Luo, Q.; Zhu, H. Key Issues and Countermeasures of Machine Vision for Fruit and Vegetable Picking Robot. Mechatronics and Automation Technology; IOS Press: Amsterdam, The Netherlands, 2024; pp. 69-78. [DOI: https://dx.doi.org/10.3233/ATDE231092]

8. Wang, C.; Pan, W.; Zou, T.; Li, C.; Han, Q.; Wang, H.; Yang, J.; Zou, X. A Review of Perception Technologies for Berry Fruit-Picking Robots: Advantages, Disadvantages, Challenges, and Prospects. Agriculture; 2024; 14, 1346. [DOI: https://dx.doi.org/10.3390/agriculture14081346]

9. Li, W.; Yin, H.; Li, Y.; Liu, X.; Liu, J.; Wang, H. Research on the Jet Distance Enhancement Device for Blueberry Harvesting Robots Based on the Dual-Ring Model. Agriculture; 2024; 14, 1563. [DOI: https://dx.doi.org/10.3390/agriculture14091563]

10. Chen, Z.; Lei, X.; Yuan, Q.; Qi, Y.; Ma, Z.; Qian, S.; Lyu, X. Key Technologies for Autonomous Fruit-and Vegetable-Picking Robots: A Review. Agronomy; 2024; 14, 2233. [DOI: https://dx.doi.org/10.3390/agronomy14102233]

11. He, F.; Zhang, Q.; Deng, G.; Li, G.; Yan, B.; Pan, D.; Luo, X.; Li, J. Research Status and Development Trend of Key Technologies for Pineapple Harvesting Equipment: A Review. Agriculture; 2024; 14, 975. [DOI: https://dx.doi.org/10.3390/agriculture14070975]

12. Zhang, Q.; Su, W.H. Real-Time Recognition and Localization of Apples for Robotic Picking Based on Structural Light and Deep Learning. Smart Cities; 2023; 6, pp. 3393-3410. [DOI: https://dx.doi.org/10.3390/smartcities6060150]

13. Gu, B.; Wen, C.; Liu, X.; Hou, Y.; Hu, Y.; Su, H. Improved YOLOv7-tiny complex environment citrus detection based on lightweighting. Agronomy; 2023; 13, 2667. [DOI: https://dx.doi.org/10.3390/agronomy13112667]

14. Yu, Y.; Liu, Y.; Li, Y.; Xu, C.; Li, Y. Object Detection Algorithm for Citrus Fruits Based on Improved YOLOv5 Model. Agriculture; 2024; 14, 1798. [DOI: https://dx.doi.org/10.3390/agriculture14101798]

15. Yan, B.; Li, X. RGB-D Camera and Fractal-Geometry-Based Maximum Diameter Estimation Method of Apples for Robot Intelligent Selective Graded Harvesting. Fractal Fract.; 2024; 8, 649. [DOI: https://dx.doi.org/10.3390/fractalfract8110649]

16. Fan, P.; Zheng, C.; Sun, J.; Chen, D.; Lang, G.; Li, Y. Enhanced Real-Time Target Detection for Picking Robots Using Lightweight CenterNet in Complex Orchard Environments. Agriculture; 2024; 14, 1059. [DOI: https://dx.doi.org/10.3390/agriculture14071059]

17. Liu, J.; Zhao, G.; Liu, S.; Liu, Y.; Yang, H.; Sun, J.; Yan, Y.; Fan, G.; Wang, J.; Zhang, H. New Progress in Intelligent Picking: Online Detection of Apple Maturity and Fruit Diameter Based on Machine Vision. Agronomy; 2024; 14, 721. [DOI: https://dx.doi.org/10.3390/agronomy14040721]

18. Wang, W.; Shi, Y.; Liu, W.; Che, Z. An Unstructured Orchard Grape Detection Method Utilizing YOLOv5s. Agriculture; 2024; 14, 262. [DOI: https://dx.doi.org/10.3390/agriculture14020262]

19. Zhao, J.; Yao, X.; Wang, Y.; Yi, Z.; Xie, Y.; Zhou, X. Lightweight-Improved YOLOv5s Model for Grape Fruit and Stem Recognition. Agriculture; 2024; 14, 774. [DOI: https://dx.doi.org/10.3390/agriculture14050774]

20. Wu, M.; Lin, H.; Shi, X.; Zhu, S.; Zheng, B. MTS-YOLO: A Multi-Task Lightweight and Efficient Model for Tomato Fruit Bunch Maturity and Stem Detection. Horticulturae; 2024; 10, 1006. [DOI: https://dx.doi.org/10.3390/horticulturae10091006]

21. Cai, Y.; Cui, B.; Deng, H.; Zeng, Z.; Wang, Q.; Lu, D.; Cui, Y.; Tian, Y. Cherry Tomato Detection for Harvesting Using Multimodal Perception and an Improved YOLOv7-Tiny Neural Network. Agronomy; 2024; 14, 2320. [DOI: https://dx.doi.org/10.3390/agronomy14102320]

22. Cao, H.; Zhang, G.; Zhao, A.; Wang, Q.; Zou, X.; Wang, H. YOLOv8n-CSE: A Model for Detecting Litchi in Nighttime Environments. Agronomy; 2024; 14, 1924. [DOI: https://dx.doi.org/10.3390/agronomy14091924]

23. Zhang, G.; Yang, X.; Lv, D.; Zhao, Y.; Liu, P. YOLOv8n-CSD: A Lightweight Detection Method for Nectarines in Complex Environments. Agronomy; 2024; 14, 2427. [DOI: https://dx.doi.org/10.3390/agronomy14102427]

24. Yang, Y.; Su, L.; Zong, A.; Tao, W.; Xu, X.; Chai, Y.; Mu, W. A New Kiwi Fruit Detection Algorithm Based on an Improved Lightweight Network. Agriculture; 2024; 14, 1823. [DOI: https://dx.doi.org/10.3390/agriculture14101823]

25. Sun, Q.; Li, P.; He, C.; Song, Q.; Chen, J.; Kong, X.; Luo, Z. A Lightweight and High-Precision Passion Fruit YOLO Detection Model for Deployment in Embedded Devices. Sensors; 2024; 24, 4942. [DOI: https://dx.doi.org/10.3390/s24154942]

26. Li, Y.; Liao, J.; Wang, J.; Luo, Y.; Lan, Y. Prototype Network for Predicting Occluded Picking Position Based on Lychee Phenotypic Features. Agronomy; 2023; 13, 2435. [DOI: https://dx.doi.org/10.3390/agronomy13092435]

27. Gonzalez, S.; Arellano, C.; Tapia, J.E. Deepblueberry: Quantification of blueberries in the wild using instance segmentation. IEEE Access; 2019; 7, pp. 105776-105788. [DOI: https://dx.doi.org/10.1109/ACCESS.2019.2933062]

28. Ni, X.; Li, C.; Jiang, H.; Takeda, F. Deep learning image segmentation and extraction of blueberry fruit traits associated with harvestability and yield. Hortic. Res.; 2020; 7, 110. [DOI: https://dx.doi.org/10.1038/s41438-020-0323-3] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32637138]

29. Luo, L.; Liu, W.; Lu, Q.; Wang, J.; Wen, W.; Yan, D.; Tang, Y. Grape berry detection and size measurement based on edge image processing and geometric morphology. Machines; 2021; 9, 233. [DOI: https://dx.doi.org/10.3390/machines9100233]

30. Chen, Y.; Li, X.; Jia, M.; Li, J.; Hu, T.; Luo, J. Instance Segmentation and Number Counting of Grape Berry Images Based on Deep Learning. Appl. Sci.; 2023; 13, 6751. [DOI: https://dx.doi.org/10.3390/app13116751]

31. Wang, Y.; Lv, J.; Xu, L.; Gu, Y.; Zou, L.; Ma, Z. A segmentation method for waxberry image under orchard environment. Sci. Hortic.; 2020; 266, 109309. [DOI: https://dx.doi.org/10.1016/j.scienta.2020.109309]

32. Pérez-Borrero, I.; Marín-Santos, D.; Gegúndez-Arias, M.E.; Cortés-Ancos, E. A fast and accurate deep learning method for strawberry instance segmentation. Comput. Electron. Agric.; 2020; 178, 105736. [DOI: https://dx.doi.org/10.1016/j.compag.2020.105736]

33. Cai, C.; Tan, J.; Zhang, P.; Ye, Y.; Zhang, J. Determining strawberries’ varying maturity levels by utilizing image segmentation methods of improved deeplabv3+. Agronomy; 2022; 12, 1875. [DOI: https://dx.doi.org/10.3390/agronomy12081875]

34. Xu, P.; Fang, N.; Liu, N.; Lin, F.; Yang, S.; Ning, J. Visual recognition of cherry tomatoes in plant factory based on improved deep instance segmentation. Comput. Electron. Agric.; 2022; 197, 106991. [DOI: https://dx.doi.org/10.1016/j.compag.2022.106991]

35. Huang, X.; Peng, D.; Qi, H.; Zhou, L.; Zhang, C. Detection and Instance Segmentation of Grape Clusters in Orchard Environments Using an Improved Mask R-CNN Model. Agriculture; 2024; 14, 918. [DOI: https://dx.doi.org/10.3390/agriculture14060918]

36. Yan, B.; Liu, Y.; Yan, W. A novel fusion perception algorithm of tree branch/trunk and apple for harvesting robot based on improved yolov8s. Agronomy; 2024; 14, 1895. [DOI: https://dx.doi.org/10.3390/agronomy14091895]

37. Santos, T.T.; De Souza, L.L.; dos Santos, A.A.; Avila, S. Grape detection, segmentation, and tracking using deep neural networks and three-dimensional association. Comput. Electron. Agric.; 2020; 170, 105247. [DOI: https://dx.doi.org/10.1016/j.compag.2020.105247]

38. Jia, W.; Zhang, Z.; Shao, W.; Ji, Z.; Hou, S. RS-Net: Robust segmentation of green overlapped apples. Precis. Agric.; 2022; 23, pp. 492-513. [DOI: https://dx.doi.org/10.1007/s11119-021-09846-3]

39. Fu, L.; Feng, Y.; Majeed, Y.; Zhang, X.; Zhang, J.; Karkee, M.; Zhang, Q. Kiwifruit detection in field images using Faster R-CNN with ZFNet. IFAC-PapersOnLine; 2018; 51, pp. 45-50. [DOI: https://dx.doi.org/10.1016/j.ifacol.2018.08.059]

40. Yang, C.; Xiong, L.; Wang, Z.; Wang, Y.; Shi, G.; Kuremot, T.; Zhao, W.; Yang, Y. Integrated detection of citrus fruits and branches using a convolutional neural network. Comput. Electron. Agric.; 2020; 174, 105469. [DOI: https://dx.doi.org/10.1016/j.compag.2020.105469]

41. Wang, C.; Ding, F.; Wang, Y.; Wu, R.; Yao, X.; Jiang, C.; Ling, L. Real-Time Detection and Instance Segmentation of Strawberry in Unstructured Environment. Comput. Mater. Contin.; 2024; 78, pp. 1481-1501. [DOI: https://dx.doi.org/10.32604/cmc.2023.046876]

42. Li, H.; Li, J.; Wei, H.; Liu, Z.; Zhan, Z.; Ren, Q. Slim-neck by GSConv: A better design paradigm of detector architectures for autonomous vehicles. arXiv; 2022; arXiv: 2206.02424[DOI: https://dx.doi.org/10.1007/s11554-024-01436-6]

43. Chen, C.; Guo, Z.; Zeng, H.; Xiong, P.; Dong, J. Repghost: A hardware-efficient ghost module via re-parameterization. arXiv; 2022; arXiv: 2211.06088[DOI: https://dx.doi.org/10.48550/arXiv.2211.06088]

44. Gou, J.; Yu, B.; Maybank, S.J.; Tao, D. Knowledge distillation: A survey. Int. J. Comput. Vis.; 2021; 129, pp. 1789-1819. [DOI: https://dx.doi.org/10.1007/s11263-021-01453-z]

45. Lin, T.Y.; Maire, M.; Belongie, S.; Hays, J.; Perona, P.; Ramanan, D.; Dollár, P.; Zitnick, C.L. Microsoft coco: Common objects in context. Proceedings of the Computer Vision–ECCV 2014: 13th European Conference; Zurich, Switzerland, 6–12 September 2014; Proceedings, Part V 13 Springer: Berlin, Germany, 2014; pp. 740-755.

46. Zhao, Y.; Lv, W.; Xu, S.; Wei, J.; Wang, G.; Dang, Q.; Liu, Y.; Chen, J. DETRs Beat YOLOs on Real-time Object Detection. Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); Seattle, WA, USA, 16–22 June 2024; pp. 16965-16974. [DOI: https://dx.doi.org/10.1109/CVPR52733.2024.01605]

47. Wang, A.; Chen, H.; Liu, L.; Chen, K.; Lin, Z.; Han, J.; Ding, G. YOLOv10: Real-Time End-to-End Object Detection. arXiv; 2024; arXiv: 2405.14458[DOI: https://dx.doi.org/10.48550/arXiv.2405.14458]

48. Ultralytics. YOLOv11: Real-Time Object Detection and Tracking. 2025; Available online: https://github.com/ultralytics/ultralytics (accessed on 1 October 2024).

49. Bolya, D.; Zhou, C.; Xiao, F.; Lee, Y.J. Yolact: Real-time instance segmentation. Proceedings of the IEEE/CVF International Conference on Computer Vision; Seoul, Republic of Korea, 27 October–2 November 2019; pp. 9157-9166. [DOI: https://dx.doi.org/10.1109/ICCV.2019.00925]

50. He, K.; Gkioxari, G.; Dollár, P.; Girshick, R. Mask r-cnn. Proceedings of the IEEE International Conference on Computer Vision; Venice, Italy, 22–29 October 2017; pp. 2961-2969. [DOI: https://dx.doi.org/10.1109/ICCV.2017.322]

51. Cheng, T.; Wang, X.; Chen, S.; Zhang, W.; Zhang, Q.; Huang, C.; Zhang, Z.; Liu, W. Sparse Instance Activation for Real-Time Instance Segmentation. Proceedings of the IEEE Conference Computer Vision and Pattern Recognition (CVPR); New Orleans, LA, USA, 18–24 June 2022.

52. Howard, A.; Sandler, M.; Chu, G.; Chen, L.C.; Chen, B.; Tan, M.; Wang, W.; Zhu, Y.; Pang, R.; Vasudevan, V. et al. Searching for mobilenetv3. Proceedings of the IEEE/CVF International Conference on Computer Vision; Seoul, Republic of Korea, 27 October–2 November 2019; pp. 1314-1324. [DOI: https://dx.doi.org/10.48550/arXiv.1905.02244]