Automated Dead Chicken Detection in Poultry Farms Using Knowledge Distillation and Vision Transformers

Abstract

Detecting dead chickens in broiler farms is critical for maintaining animal welfare and preventing disease outbreaks. This study presents an automated system that leverages CCTV footage to detect dead chickens, utilizing a two-step approach to improve detection accuracy and efficiency. First, stationary regions in the footage—likely representing dead chickens—are identified. Then, a deep learning classifier, enhanced through knowledge distillation, confirms whether the detected stationary object is indeed a chicken. EfficientNet-B0 is employed as the teacher model, while DeiT-Tiny functions as the student model, balancing high accuracy and computational efficiency. A dynamic frame selection strategy optimizes resource usage by adjusting monitoring intervals based on the chickens’ age, ensuring real-time performance in resource-constrained environments. This method addresses key challenges such as the lack of explicit annotations for dead chickens, along with common farm issues like lighting variations, occlusions, cluttered backgrounds, chicken growth, and camera distortions. The experimental results demonstrate validation accuracies of 99.3% for the teacher model and 98.7% for the student model, with significant reductions in computational demands. The system’s robustness and scalability make it suitable for large-scale farm deployment, minimizing the need for labor-intensive manual inspections. Future work will explore integrating deep learning methods that incorporate temporal attention mechanisms and automated removal processes.

1. Introduction

Poultry farming, particularly large-scale broiler operations, faces the critical challenge of maintaining flock health through continuous monitoring. The early detection of dead chickens is vital to prevent contamination, disease spread, and significant economic losses [1]. Identifying and removing dead chickens is critical for maintaining flock health in broiler farms. However, traditional manual inspection methods are increasingly insufficient due to their reliance on human effort, which is labor-intensive and prone to error in large-scale operations. In high-density environments, where thousands of chickens are housed in confined spaces, manual detection becomes inefficient and inconsistent, further escalating the risk of disease outbreaks [2]. With the rising global demand for poultry, there is a growing need for scalable and automated systems to safeguard the health and welfare of broiler flocks [3].

Existing automated detection systems typically rely on classical computer vision techniques or traditional machine learning models such as Random Forests (RFs) [4] and Support Vector Machines (SVMs) [5,6]. These methods classify chickens based on manually extracted features like color, texture, or shape. However, in real-world farm environments, these models struggle to maintain accuracy due to dynamic challenges such as fluctuating lighting, occlusions, cluttered backgrounds, and the continuous growth of chickens [7]. Additionally, distinguishing between dead and stationary chickens becomes difficult, especially in high-density settings where visual similarities exist. These conventional methods struggle to maintain consistent performance, particularly when faced with the diverse and unpredictable conditions of modern broiler farms.

Advancements in deep learning have opened new possibilities for automating tasks in agriculture. Deep learning models, particularly convolutional neural networks (CNNs) [8] and vision transformers (ViTs) [9], have demonstrated significant improvements in object detection and classification across various domains, including crop monitoring and livestock analysis [3,7,10,11,12,13,14,15,16,17]. In this context, knowledge distillation has emerged as an effective technique to train smaller, efficient models (students) by transferring knowledge from larger, more accurate models (teachers). This process helps reduce the computational complexity of the student model without compromising accuracy, making it ideal for applications in real-time, resource-constrained environments like broiler farms [18].

A major challenge in training deep learning models for dead chicken detection is the absence of explicit annotations for dead chickens, making it difficult to create accurately labeled datasets. Manually labeling dead and stationary chickens is labor-intensive and prone to mistakes due to their visual similarities. To address this, our approach first detects stationary regions likely to represent dead chickens and then applies a deep learning classifier to confirm whether the detected object is a chicken. This two-stage method leverages EfficientNet [19] as the teacher model and a distilled vision transformer (DeiT) [20] as the student, helping to overcome the lack of precise annotations while maintaining high performance in real-world farm conditions.

Our proposed approach addresses the limitations of existing methods by integrating stationary object detection with a knowledge-distilled classifier, achieving superior accuracy compared to traditional models like Random Forest. The two-stage method effectively distinguishes dead from inactive chickens, ensuring robustness in high-density settings. Knowledge distillation enables real-time applicability by creating a lightweight yet powerful student model, while the dynamic frame capture strategy improves processing efficiency, making the system scalable for large-scale broiler farms.

This paper presents a novel automated system for detecting dead chickens using CCTV footage, combining deep learning techniques and knowledge distillation to enhance detection accuracy and efficiency. The system integrates stationary object detection with EfficientNet and DeiT for classification and employs a dynamic frame selection strategy that adjusts the monitoring frequency based on the chickens’ age. The key contributions of this work include the following:

• A two-stage approach for detecting dead chickens that combines stationary object detection with a deep learning classifier, compensating for the absence of explicit dead chicken annotations by filtering out temporary inactivity and confirming that stationary objects are chickens.

• The use of knowledge distillation, where EfficientNet acts as the teacher model and DeiT as the student, ensuring a balance between high-quality feature extraction and computational efficiency, enabling real-time deployment in resource-constrained farm environments.

• A dynamic frame capture strategy, based on chicken age, which optimizes the trade-off between computational load and detection accuracy, crucial for scalable deployment in large-scale farms.

• Extensive experiments demonstrating superior accuracy and robustness, showcasing the system’s ability to manage real-world challenges such as illumination changes, occlusions, and cluttered environments.

2. Related Work

Historically, manual inspection has been the predominant approach for detecting dead chickens on broiler farms, requiring no additional tools or technology. However, its effectiveness diminishes in high-density farming environments, where the sheer scale of operations and potential for human oversight make it unsustainable [2,17]. Previous attempts to automate the process have relied on machine learning models such as Random Forests (RFs) [4] and Support Vector Machines (SVMs) [5,6,21], which classify dead chickens based on manually extracted features like color, texture, or shape. Zhu et al. [22] used an SVM for dead chicken detection. Pirompud et al. [23] employed RF to predict dead-on-arrival incidents of broiler chickens. Bao et al. [24] proposed a dead and sick chicken detection method by fastening a foot ring on each chicken to capture the activity intensity of the chicken and using a machine learning classification method to identify the status of the chicken. Existing models often fail to generalize effectively in real-world farm environments, where challenges such as varying lighting, occlusions, and crowded spaces significantly impact their accuracy [12].

Recent advancements in deep learning have led to significant improvements in object detection and classification in agriculture, particularly in crop monitoring and livestock behavior analysis [12]. Liu et al. [15] used YOLOv4 [25] for the identification of dead chickens. Bist et al. [16] used YOLOv5s for automated early mortality detection in a cage-free environment. Yang et al. [14] proposed improved YOLOv7 [26] for detecting dead caged hens. Transformers, originally developed for natural language processing, have recently gained traction in computer vision tasks [9]. Luo et al. [13] used Deformable DETR for the detection of dead laying hens on a commercial farm. ViTs including DeiT are increasingly being used for image classification due to their ability to capture long-range dependencies in data.

Dynamic frame capture strategies are crucial in surveillance systems where object appearance and movement change over time. By filtering frames, these strategies reduce storage and processing complexities, improving efficiency and ensuring critical events are captured without overwhelming the system [27].

Knowledge distillation, where a smaller, more efficient student model is trained using the knowledge learned by a larger teacher model, has been shown to reduce computational complexity while maintaining high accuracy [20]. For example, Huang et al. [28] developed a multistage knowledge distillation model for lightweight and efficient plant disease detection. Cai et al. [29] used knowledge distillation-based methods for the real-time semantic segmentation of surface defects in navel oranges. Tong et al. [11] proposed a similar scheme for enhancing chicken health detection.

One key challenge in deploying deep learning models for dead chicken detection is the lack of high-quality, labeled datasets. Most broiler farm datasets do not explicitly annotate dead chickens, and visually distinguishing between stationary and dead chickens is difficult due to their similarity. Creating fully annotated datasets is resource-intensive, prone to human error, and impractical in large-scale farm operations.

To address this, we propose a two-step approach. First, stationary regions in the footage—likely representing dead chickens—are detected. Then, a deep learning classifier, leveraging knowledge distillation, is applied to confirm that the detected stationary object is a chicken, refining the classification. By employing EfficientNet-B0 as the teacher model and DeiT-Tiny as the student, the knowledge distillation process enables the student model to generalize effectively, even with limited annotated data. Combined with dynamic frame capture, this method balances real-time performance with computational efficiency, making it ideal for high-density broiler farm environments.

3. Method

3.1. Model Architecture

The proposed model architecture consists of two key components: a stationary object detection module and a classification module, which work together to detect dead chickens in CCTV footage accurately. This two-stage approach is particularly effective given the lack of annotated data for dead chickens, which makes direct deep learning training difficult. To reduce computational overhead, traditional image processing techniques are first employed to detect stationary regions that are likely to indicate dead chickens.

However, not all stationary objects are dead chickens—some may be inanimate objects, leading to potential false positives. To address this, the second stage incorporates a lightweight knowledge-distilled vision transformer classifier that further analyzes the detected objects. This step ensures accurate differentiation between chickens and non-chicken objects, significantly improving detection precision.

As shown in Figure 1, frames are dynamically selected from the CCTV footage based on the age of the chickens. These frames are then processed by the stationary object detection module to identify regions that remain stationary across several consecutive frames. It can effectively differentiate between inactive and dead chickens. Inactive chickens will show slight movements even during rest, and dead chickens will not exhibit any movement. The dynamic interval accounts for the natural behavior of chickens, ensuring that only objects that do not move are flagged for further analysis. To minimize false positives, detected objects are filtered by size, aligning with the expected size of the chickens at their specific age, thus discarding irrelevant objects. In the second stage, the filtered objects are classified using the knowledge-distilled vision transformer model, which determines whether the object is a chicken. If confirmed as a chicken and observed to remain stationary across multiple frames, it is classified as a dead chicken.

3.1.1. Stationary Object Detection

The proposed system detects stationary white areas within the frames over time by employing a series of image processing steps to identify potential dead chickens. As shown in Figure 2, the process begins with frame differencing, where consecutive frames are compared to detect regions with minimal or no motion. This allows the system to isolate stationary areas that are likely to indicate inactive chickens. Next, binarization is applied to the detected differences, converting the grayscale differences into binary masks, which highlight stationary regions within the frame. Contour extraction is then performed on the binary masks, delineating the regions of interest precisely by outlining the stationary areas that could potentially represent dead chickens. To further refine the detection, white masking is applied to regions that remain stationary over multiple frames. This step is crucial in filtering out transient pauses in chicken movement, such as when a chicken temporarily rests, ensuring that only consistently stationary regions are flagged for further analysis. By focusing on stationary regions over time and combining these multiple stages of image processing, the system effectively reduces false positives and enhances the overall accuracy and reliability of dead chicken detection. These regions are then passed to the classification stage for further verification.

3.1.2. Chicken Classifier

After identifying the stationary regions, the classifier determines whether these areas represent chickens by leveraging a combination of deep learning models in a knowledge distillation framework. The EfficientNet [19] model is employed as the teacher model. Pre-trained on ImageNet, EfficientNet-B0 is fine-tuned using a custom dataset containing images of chickens and non-chicken objects captured from the farm environment. It is selected due to its high accuracy, efficiency, and ability to extract detailed features from the input images, providing a robust foundation for training the student model. Its strong performance in feature extraction makes it an ideal teacher model for knowledge distillation.

The distilled vision transformer model DeiT-Tiny [20] is used as the student model. DeiT (distilled vision transformer) is a lightweight version of the vision transformer (ViT) designed for resource-constrained environments. Unlike traditional convolutional neural networks, ViT models operate on sequences of image patches, leveraging self-attention mechanisms to capture long-range dependencies in the input data. DeiT improves upon ViT by incorporating a knowledge distillation process during training, allowing it to achieve high accuracy with significantly reduced computational complexity.

In DeiT, an additional distillation token is used to directly learn from the teacher model during training, enhancing the student model’s performance while retaining its efficiency. The use of a lightweight architecture, combined with the distillation process, enables DeiT-Tiny to deliver competitive classification accuracy with minimal resource requirements, making it suitable for real-time applications in broiler farms.

Through the knowledge distillation process [18], which employs a distillation loss function, the model learns to replicate the intricate feature extraction capabilities of the more complex teacher model. This process enables the student model to achieve high accuracy while maintaining a lightweight structure, which is essential for real-time inference, particularly in resource-limited environments like farms.

By combining these models, the system ensures that the student model retains the classification accuracy and robust feature extraction of the teacher model while also being computationally efficient enough to perform the real-time detection of dead chickens with minimal hardware requirements. The accuracy of the teacher and the lightweight structure of the student make this combination highly effective for practical deployment in broiler farms.

Knowledge distillation is a process in which a lightweight student model is trained to mimic the performance of a larger, more complex teacher model. In this approach, the student model is optimized using both soft targets—the probability distribution outputs from the teacher model—and hard targets, which are the true labels from the dataset. The soft targets provide rich information about the relationships between classes, while the hard targets offer direct supervision. To integrate these two types of learning, a distillation loss function is employed, combining the standard classification loss of the student model with the distillation loss from the teacher model [18]. The overall loss function (L) can be expressed as follows:

L = α × L_CE + (1 − α) × L_KD(1)

where L_CE is the cross-entropy loss between the student model’s predictions and the true labels (hard targets). L_KD is the distillation loss, calculated as the Kullback–Leibler divergence between the teacher and student outputs (soft targets), often modulated by a temperature parameter T to soften the probabilities. α is a weighting factor balancing the importance of the classification loss and the distillation loss.

A temperature parameter T is often applied to smooth the teacher’s output probabilities, which helps the student model effectively learn from the teacher. Typically, T is set to a value greater than 1, softening the probabilities and making it easier for the student model to absorb the nuanced knowledge of the teacher model.

By leveraging knowledge distillation, the student model retains much of the teacher model’s performance while being significantly more computationally efficient [30]. This makes it well suited for real-time inference in resource-constrained environments like broiler farms. The technique ensures high classification accuracy and enables the practical deployment of deep learning models in settings with limited computational capacity.

In chicken farms, knowledge distillation enhances student model performance for the real-time detection of dead chickens with reduced resource requirements. This approach streamlines farm operations by minimizing the need for manual inspections and enabling more scalable, efficient monitoring systems that promote better animal welfare and management.

3.2. Frame Selection

To ensure efficient processing and accurate detection, the system utilizes a dynamic frame selection strategy tailored to the age of the chickens, based on video footage from CCTV cameras installed in broiler farms. This approach balances processing load and detection accuracy by adapting to the changing activity levels of chickens as they age.

As shown in Figure 3, for younger chickens, less than 7 days old, frames are captured every 3 s to account for their higher activity levels and rapid movements. As the chickens age, the intervals are progressively shortened: for chickens between 7 and 14 days old, frames are captured every 2 s; for those between 14 and 21 days old, every 1 s; and for chickens between 21 and 28 days old, every 0.5 s. Finally, for chickens older than 28 days, frames are captured every 0.25 s to effectively monitor prolonged periods of inactivity, which could indicate the presence of dead chickens. These intervals were determined through experimental testing and interpolation, ensuring that the system remains responsive to behavioral changes, optimizing both detection accuracy and processing efficiency throughout the chickens’ growth stages.

The dynamic frame capture strategy significantly improves computational efficiency, allowing the system to reduce the number of frames processed without sacrificing detection accuracy. By adapting frame capture intervals based on the age of the chickens, the system optimizes resource usage, which is particularly important in large-scale farm operations where resource management is critical. This balance ensures that computational resources are allocated efficiently, monitoring more active younger chickens less frequently while closely observing older, more inactive chickens.

3.3. Object Size Filtering

To enhance detection accuracy and minimize false positives, the system employs an object size filtering strategy based on the expected size of chickens at various growth stages. Detected contours (stationary objects) from the CCTV footage are evaluated to ensure they align with the anticipated dimensions of the chickens. Through the empirical analysis of the dataset, a minimum area threshold of 200 pixels and a maximum area threshold of 800 pixels were determined. These thresholds correspond to the approximate sizes of chickens across different growth stages, taking into account variations in posture and viewing angles observed in the footage.

The selection of these parameters was guided by extensive experimentation using video frames. Frames were analyzed to identify the pixel dimensions of chickens at different stages of growth, and the thresholds were fine-tuned to minimize false positives. Additionally, aspect ratios were evaluated to exclude irregular shapes that do not resemble a chicken’s general form. Solidity metrics were also employed to filter out contours with jagged edges or hollow interiors, further refining the detection process.

By incorporating these empirically derived size and shape criteria, the system ensures that only valid detections proceed to the classification stage, where the knowledge-distilled vision transformer model confirms whether the object is indeed a chicken. This targeted approach not only boosts detection precision but also optimizes computational resources by focusing processing efforts on relevant objects, crucial for efficient monitoring in large-scale broiler farm operations.

3.4. Training Dataset Preparation

The dataset comprises video footage from a broiler farm, capturing chickens at various growth stages and environmental conditions. The annotations only mark the presence of chickens, with no explicit labels for dead chickens. To address this, a two-step detection approach was employed, focusing on identifying stationary regions within the footage as potential indicators of dead chickens. Additionally, footage of the empty farm was used to generate images of non-chicken objects for negative samples, enhancing the classification process. The dataset encompasses many scenarios, including densely packed chickens, fluctuating illumination, and different environmental factors, ensuring the detection model can generalize effectively across different real-world conditions.

3.5. Source of the Dataset

The video footage for this study was captured at Daseong Farm, located in Namwon-si, Republic of Korea, using an IPC675LFW-AX4DUPKC-VG camera. Weekly recordings were conducted to document the progression of chicken growth under diverse environmental conditions, such as high-density flock arrangements and varying lighting levels.

For the training dataset to the classifier, positive samples were extracted by cropping chickens from the frames, ensuring a wide variety of conditions were represented. These samples included chickens across all growth stages, from small chicks to fully grown birds, and covered different flock densities, ranging from tightly packed groups to more dispersed arrangements. Additionally, the dataset accounted for various lighting conditions and dim or shadowed regions, simulating the challenges typically encountered in real-world farm environments.

Negative samples were derived from frames that did not contain chickens, captured from footage of the empty farm. Non-chicken objects in the farm environment, such as equipment and background elements, were also included to enhance the classifier’s ability to distinguish chickens from other objects. These negative samples helped reduce the likelihood of false positives and improved the classifier’s generalization capabilities.

The dataset comprises 7062 positive samples (chickens) and 7000 negative samples (non-chickens). It was divided into training (70%), validation (15%), and test (15%) sets, ensuring a balanced representation of various growth stages, flock densities, and environmental conditions across each split. This division ensured the model would generalize well to unseen data, leading to robust performance in practical farm environments.

3.6. Experimental Setup

The experiments were conducted on a system equipped with an NVIDIA RTX 3090 GPU and 64 GB of RAM. The software environment included PyTorch 1.9 and Python 3.8. The teacher and the student model were trained using the AdamW optimizer and cross-entropy loss, with an additional distillation loss applied to the student model. The training process was configured with a batch size of 32 and a learning rate of 0.001 and ran for 50 epochs.

3.7. Evaluation Metrics

The model was evaluated using accuracy, precision, recall, F1-score, and average precision (AP). Accuracy measures the ratio of correctly classified instances to the total instances. Precision evaluates the accuracy of the positive predictions. It evaluates the ratio of true positive predictions to the total positive predictions. Recall measures the model’s ability to identify true positive cases. It is the ratio of true positive predictions to the actual positive instances. The F1-score combines precision and recall into a single metric by calculating their harmonic mean. The AP is the average of precision values at various levels of recall. These metrics are formulated as follows:

(2) $Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$

(3) $Precision={\displaystyle \frac{TP}{TP+FP}}$

(4) $Recall={\displaystyle \frac{TP}{TP+FN}}$

(5) $F1-score=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}}$

(6)$AP={{\displaystyle \sum }}_{n}P\left(R_n\right)\Delta R_n$

where TP is true positive and represents chickens correctly classified, TN is true negative and refers to non-chickens correctly classified, FP is false positive which denotes non-chickens incorrectly classified as chickens, FN is false negative and represents chickens incorrectly classified as non-chickens, P is the precision, R is the recall, $P\left(R_n\right)$ is the precision at the recall level $R_n$, and $\Delta R_n$ is the change in recall from the previous recall level to $R_n$.

4. Experiments and Results

4.1. Training and Validation Performance of the Classifier

The training and validation performance of both the teacher (EfficientNet-B0) and student (DeiT-Tiny) models were precisely tracked for 50 epochs. The training loss consistently decreased, indicating effective learning, while the validation loss plateaued after approximately 30 epochs, suggesting convergence. Both models exhibited a steady increase in training accuracy, reaching up to 98% towards the final epochs. Similarly, the validation accuracy showed substantial improvement, stabilizing around 99% in the later stages of training. Notably, the teacher model demonstrated a slightly higher convergence speed than the student model, attributed to its more complex architecture.

In Figure 4, the loss curves illustrate the decline in training and validation loss over the epochs for both the teacher and student models. This indicates that both models effectively minimize classification errors during training. Similarly, the accuracy curve depicts the improvement in training and validation accuracy, highlighting the models’ ability to correctly classify chickens and non-chickens with high precisions.

4.2. Classification Performance

The final classification performance metrics for both models were evaluated on the validation and test datasets. The teacher model achieved a validation accuracy of 99.3%, a precision of 98.8%, a recall of 99.1%, and an F1-score of 98.9%. The student model, leveraging knowledge distillation, attained a validation accuracy of 98.7%, a precision of 98.2%, a recall of 98.6%, and an F1-score of 98.4%. On the test dataset, the teacher model maintained slightly superior performance metrics compared to the student model. The confusion matrices for both datasets reveal high true positive and true negative rates, with minimal false positives and false negatives, underscoring the models’ robustness in classification tasks.

Table 1 presents the final classification metrics for both the teacher and student models across the validation and test datasets. The teacher model consistently outperformed the student model, albeit with a marginal difference, showcasing the effectiveness of knowledge distillation in maintaining high accuracy with a lightweight architecture. Figure 5 displays the confusion matrices for the teacher and student models on the validation dataset. Both models demonstrated high true positives referring to chickens correctly identified and higher true negatives stating the rates of non-chickens correctly identified, with negligible misclassifications.

4.3. Teacher and Student Model Comparison

The comparison between the teacher and the student model highlights the effectiveness of knowledge distillation in balancing performance and computational efficiency. The teacher model achieved a 99.3% validation accuracy, while the student model reached 98.7%. Despite the slight performance difference, the student model offered a significant advantage in inference time and model size, with 30% faster inference and 40% smaller memory footprint.

Table 2 summarizes the comparative performance and computational metrics of the EfficientNet-B0 teacher model and the DeiT-Tiny student model. While the teacher model exhibited slightly higher accuracy, the student model achieved nearly comparable performance with a notable reduction in inference time and model size. Figure 6 visually represents these efficiency gains, emphasizing the student model’s suitability for real-time deployment in resource-constrained environments like chicken farms.

4.4. Student Model Evaluation

To further optimize the classification module, three DeiT architectures, DeiT-Tiny, DeiT-Small, and DeiT-Base, were evaluated. Table 3 shows the performance metrics for each model.

While DeiT-Base offered the highest validation accuracy and average precision, it has a significantly larger model size and longer inference time, which may not be ideal for real-time deployment in resource-constrained environments. DeiT-Tiny, on the other hand, provided a commendable validation accuracy of 98.7% and an average precision of 0.990 with the lowest inference time and smallest model size. DeiT-Small presented a middle ground but still falls short in computational efficiency compared to DeiT-Tiny.

Given the marginal performance gains of DeiT-Base and DeiT-Small over DeiT-Tiny, coupled with their increased computational demands, DeiT-Tiny was selected as the student model. This choice ensures a balance between maintaining high classification accuracy and achieving the necessary computational efficiency for real-time deployment in broiler farms.

4.5. Teacher Model Evaluation

To identify the most effective teacher model for knowledge distillation, several architectures were evaluated. The performance metrics for each teacher model are summarized in Table 4.

From Table 4, EfficientNet-B0 demonstrated the highest balance between accuracy, inference time, and model size among the evaluated teacher models. Although models like ResNet-152 achieved comparable accuracy, their significantly larger size and longer inference times make them less suitable for resource-constrained environments. EfficientNet-B0’s superior performance and efficiency make it the optimal choice for our knowledge distillation framework.

4.6. Qualitative Analysis

For the qualitative assessment of the models’ performance, several examples of correct detections and common failure cases were reviewed. The correct detections involved accurately identifying chickens amidst varying backgrounds, lighting conditions, and occlusions. These successes can be attributed to the algorithm’s design: the stationary object detection module effectively identifies regions with no motion over time, while the classification module, leveraging knowledge distillation from EfficientNet-B0 to DeiT-Tiny, extracts detailed features such as texture, posture, and orientation. The dynamic frame capture intervals and object size filter further enhance this process by adapting to the age of the chickens, minimizing false positives.

Figure 7 showcases examples where the model successfully detected dead chickens under challenging conditions, including cluttered backgrounds and varying lighting. The high accuracy in these scenarios highlights the robustness of the knowledge-distilled DeiT model, which generalizes well across diverse conditions. The teacher–student distillation framework enables the DeiT model to retain the powerful feature extraction capabilities of EfficientNet-B0 while maintaining computational efficiency.

Failure cases primarily included instances where shadows or similar-looking non-chicken objects, such as feeding plates or equipment, were mistakenly classified as chickens. These errors occur due to overlapping visual characteristics, such as shape or color similarity, or due to the placement of these objects within the frame. Additionally, extreme lighting conditions occasionally impacted the classifier’s ability to extract distinguishing features, leading to false positives.

To address these challenges, further improvements are proposed, such as incorporating a temporal attention mechanism to analyze motion patterns over time and adding preprocessing steps like shadow detection and removal. These enhancements could reduce false positives by refining the distinction between chickens and non-chicken objects.

Despite these occasional misclassifications, the mistakes were infrequent, underscoring the models’ high discriminative capabilities and overall reliability. The proposed algorithm balances accuracy and computational efficiency, making it a robust solution for real-time applications in high-density broiler farm environments.

5. Discussion

The proposed method for detecting dead chickens in broiler farms utilizes a two-stage approach incorporating dynamic frame selection and knowledge distillation, resulting in significant improvements in accuracy, efficiency, and practicality compared to traditional methods. This combination effectively addressed the challenge of the absence of explicit annotations for dead chickens and the need for real-time performance in resource-constrained environments.

The dynamic frame capture strategy enhanced the computational efficiency by reducing the number of processed frames without sacrificing detection accuracy. By adapting capture intervals based on the age of the chickens, the system optimized resource usage, which is crucial in large-scale farm operations. This allows for the less frequent monitoring of active younger chickens while closely observing older, more inactive birds.

The two-stage architecture, combining stationary object detection with a knowledge-distilled vision transformer classifier, was crucial for both performance and efficiency. Given the lack of annotated data for dead chickens, traditional image processing was first used to detect stationary regions, reducing computational load. However, since not all stationary objects are dead chickens, the second stage applies a classifier to confirm whether the detected object is a chicken, improving accuracy. For the classifier, EfficientNet-B0, as the teacher model, provided robust feature extraction, while DeiT-Tiny, the student model, achieved comparable accuracy (98.7%) with lower computational demands. Knowledge distillation retained much of the teacher’s performance while optimizing the system for real-time deployment.

Performance evaluations demonstrated that EfficientNet-B0 outperformed other teacher models like ResNet-50 and EfficientNet-B1, reinforcing its selection as the teacher model. Among student models, DeiT-Tiny proved most computationally efficient, with faster inference times and smaller memory requirements than DeiT-Small and DeiT-Base.

The system’s lightweight architecture supports deployment on inexpensive hardware while maintaining high accuracy. The dynamic frame selection mechanism enhances efficiency by reducing unnecessary processing and focusing on critical frames, thus minimizing the need for labor-intensive manual inspections. This scalable approach improves overall farm management and animal welfare, making it well suited for commercial broiler farms.

6. Conclusions and Further Research

This study presents a robust and scalable system for automated dead chicken detection in broiler farms using CCTV footage. The proposed two-stage architecture, which combined stationary object detection with a knowledge-distilled vision transformer classifier, addressed the lack of annotated data for dead chickens. By employing traditional image processing to detect stationary regions first, followed by a deep learning classifier to confirm that the detected stationary object is indeed a chicken, the system significantly improved detection accuracy while maintaining computational efficiency.

The system demonstrated excellent performance, with the EfficientNet-B0 teacher model achieving 99.3% validation accuracy and the DeiT-Tiny student model achieving 98.7%. Knowledge distillation enabled the student model to retain much of the teacher’s feature extraction capabilities while reducing computational overhead, making the system well suited for real-time deployment in resource-constrained environments. The dynamic frame selection strategy further optimized resource usage by adjusting the monitoring frequency based on chicken age, ensuring efficient processing throughout the chickens’ growth stages.

The system successfully tackled real-world farm challenges, such as lighting variations, occlusions, cluttered backgrounds, chicken growth, and camera distortions, proving to be both robust and reliable. This advancement minimizes the labor-intensive task of manual inspections and enhances farm management by providing an efficient, scalable monitoring solution.

Future research will focus on expanding the dataset to include multiple farm environments, integrating a deep learning approach based on temporal attention mechanisms in video for the improved detection of subtle behavioral anomalies, and exploring automated removal systems for detected dead chickens. Temporal attention mechanisms will enable the model to analyze motion patterns across consecutive frames, improving its ability to detect behaviors and conditions indicative of chicken death. By capturing the temporal dependencies in the video data, this approach can refine the classification and reduce false positives or missed detections. These enhancements will further validate the system’s generalizability and operational effectiveness, paving the way for widespread adoption in the poultry farming industry.

Author Contributions

Conceptualization, R.K. and J.L.; data curation, R.K. and J.L.; formal analysis, R.K.; funding acquisition, J.L.; investigation, R.K.; methodology, R.K. and J.L.; resources, J.L.; software, R.K. and W.W.; validation, R.K. and W.W.; visualization, R.K. and W.W.; writing—original draft, R.K.; writing—review and editing, W.W. and J.L. All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The original contributions presented in the study are included in the article, further inquiries can be directed to the corresponding author.

Acknowledgments

We would like to express our gratitude to our lab members Nikhil Thapa and Yoochan Choi for their support in dataset curation.

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. The pipeline for the two-stage dead chicken detection.

Figure 2. Stationary object detection process.

Figure 3. Chart and block diagram showing the frame selection rate across the age of the chicken (a,b). In (b), frames from the CCTV footage are represented in blue and orange colors. The orange color represents the captured frames for evaluation.

Figure 4. Training and validation loss and accuracy curves: (a,b) are the training and validation loss curves for teacher and student models, respectively. (c,d) are the training and validation accuracy curves for both models.

Figure 5. Confusion matrix for teacher and student model on the validation dataset.

Figure 6. Bar chart comparing inference time and model size.

Figure 7. Examples of correct detections in (a,b) and common failure cases in (c). The first column contains the images of stationary object detection (within green box), and the second column contains the images of the confirmed dead chickens (within red box).

Table 1

Final classification metrics for teacher and student models on validation and test datasets.

Model	Dataset	Accuracy	Precision	Recall	F1-Score
EfficientNet-B0	Validation	99.3%	98.8%	99.1%	98.9%
	Test	99.0%	98.5%	98.8%	98.6%
DeiT-Tiny	Validation	98.7%	98.2%	98.6%	98.4%
	Test	98.5%	97.9%	98.3%	98.1%

Table 2

Performance metrics and computational efficiency comparison.

Metric	EfficientNet-B0 (Teacher)	Deit-Tiny (Student)
Validation accuracy	99.3%	98.7%
Test accuracy	99.0%	98.5%
AP	0.995	0.990
Inference time (ms/frame)	15	10
Model size (MB)	28.4	17.0

Table 3

Comparison of different DeiT models.

Metric	DeiT-Tiny	DeiT-Small	DeiT-Base
Validation accuracy	98.7%	99.0%	99.2%
Average precision (AP)	0.990	0.992	0.995
Inference time (ms/frame)	10	12	15
Model size (MB)	17.0	25.3	60.2

Table 4

Performance metrics for various teacher models.

Model	Accuracy	Precision	Recall	F1-Score	Inference Time (ms/Frame)	Model Size (MB)
EfficientNet-B0	99.3%	98.8%	99.1%	98.9%	15	28.4
EfficientNet-B1	99.1%	98.6%	98.9%	98.7%	18	34.6
ResNet-50	98.9%	98.4%	98.7%	98.5%	20	25.6
ResNet-152	99.0%	98.5%	98.8%	98.6%	25	58.2
DenseNet169	98.8%	98.2%	98.5%	98.3%	22	26.7
DenseNet201	98.7%	97.9%	98.2%	98.0%	24	32.4

References

1. Gržinić, G.; Bilandžić, N.; Škrbić, B.; Capak, D.; Velki, M.; Matoničkin Kepčija, R.; Reljić, S. Intensive Poultry Farming: A Review of the Impact on the Environment and Human Health. Sci. Total Environ.; 2023; 859, 160014. [DOI: https://dx.doi.org/10.1016/j.scitotenv.2022.160014] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36368402]

2. Nielsen, S.S.; Alvarez, J.; Bicout, D.J.; Calistri, P.; Canali, E.; Drewe, J.A.; Garin-Bastuji, B.; Gonzales Rojas, J.L.; Herskin, M.; Michel, V. et al. Welfare of Broilers on Farm. EFSA J.; 2023; 21, e07788. [DOI: https://dx.doi.org/10.2903/j.efsa.2023.7788] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36824680]

3. George, A.S.; George, A.S.H. Optimizing Poultry Production through Advanced Monitoring and Control Systems. Partn. Univers. Int. Innov. J.; 2023; 1, pp. 77-97. [DOI: https://dx.doi.org/10.5281/zenodo.10050352]

4. Breiman, L. Random Forests. Mach. Learn.; 2001; 45, pp. 5-32. [DOI: https://dx.doi.org/10.1023/A:1010933404324]

5. Hearst, M.A. Support Vector Machines. IEEE Intell. Syst.; 1998; 13, pp. 18-28. [DOI: https://dx.doi.org/10.1109/5254.708428]

6. Liu, C.; Wang, L.; Yang, A.; Zhang, Y. Support Vector Machine Classification Algorithm and Its Application. Proceedings of the 2012 International Conference on Systems and Informatics; Yantai, China, 19–20 May 2012; pp. 172-176.

7. Khanal, R.; Choi, Y.; Lee, J. Transforming Poultry Farming: A Pyramid Vision Transformer Approach for Accurate Chicken Counting in Smart Farm Environments. Sensors; 2024; 24, 2977. [DOI: https://dx.doi.org/10.3390/s24102977]

8. O’Shea, K.; Nash, R. An Introduction to Convolutional Neural Networks. arXiv; 2015; arXiv: 1511.08458v2

9. Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S. et al. An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale. arXiv; 2010; arXiv: 2010.11929v2

10. Vrindavanam, J.; Kumar, P.; Patil, G. Poultry Disease Identification in Fecal Images Using Vision Transformer. Medicon Agric. Environ. Sci.; 2023; 6, pp. 5-15. [DOI: https://dx.doi.org/10.55162/mcaes.06.150]

11. Tong, Q.; Jiang, F.; Liu, Z.; Yang, X.; Ma, Y.; Wang, H. Edge AI-Enabled Chicken Health Detection Based on Enhanced FCOS-Lite and Knowledge Distillation. Comput. Electron. Agric.; 2024; 226, 109432. [DOI: https://dx.doi.org/10.1016/j.compag.2024.109432]

12. Altalak, M.; Uddin, M.A.; Alajmi, A.; Rizg, A. Smart Agriculture Applications Using Deep Learning Technologies: A Survey. Appl. Sci.; 2022; 12, 5919. [DOI: https://dx.doi.org/10.3390/app12125919]

13. Luo, S.; Ma, Y.; Jiang, F.; Wang, H.; Tong, Q.; Wang, L. Dead Laying Hens Detection Using TIR-NIR-Depth Images and Deep Learning on a Commercial Farm. Animals; 2023; 13, 1861. [DOI: https://dx.doi.org/10.3390/ani13111861] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37889777]

14. Yang, J.; Zhang, T.; Fang, C.; Zheng, H.; Ma, C.; Wu, Z. A Detection Method for Dead Caged Hens Based on Improved YOLOv7. Comput. Electron. Agric.; 2024; 226, 109388. [DOI: https://dx.doi.org/10.1016/j.compag.2024.109388]

15. Liu, H.W.; Chen, C.H.; Tsai, Y.C.; Hsieh, K.W.; Lin, H.T. Identifying Images of Dead Chickens with a Chicken Removal System Integrated with a Deep Learning Algorithm. Sensors; 2021; 21, 3597. [DOI: https://dx.doi.org/10.3390/s21113579]

16. Bist, R.B.; Subedi, S.; Yang, X.; Chai, L. Automatic Detection of Cage-Free Dead Hens with Deep Learning Methods. AgriEngineering; 2023; 5, pp. 1020-1038. [DOI: https://dx.doi.org/10.3390/agriengineering5020064]

17. Yajie, L.; Johar, M.G.M.; Hajamydeen, A.I. Poultry Disease Early Detection Methods Using Deep Learning Technology. Indones. J. Electr. Eng. Comput. Sci.; 2023; 32, pp. 1712-1723. [DOI: https://dx.doi.org/10.11591/ijeecs.v32.i3.pp1712-1723]

18. Hinton, G.; Dean, J. Distilling the Knowledge in a Neural Network. arXiv; 2015; arXiv: 1503.02531v1

19. Tan, M.; Le, Q.V. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. arXiv; 2019; arXiv: 1905.11946v5

20. Touvron, H.; Cord, M.; Douze, M.; Massa, F.; Sablayrolles, A.; Jégou, H. Training Data-Efficient Image Transformers & Distillation Through Attention. arXiv; 2021; arXiv: 2012.12877v2

21. Zhuang, X.; Bi, M.; Guo, J.; Wu, S.; Zhang, T. Development of an Early Warning Algorithm to Detect Sick Broilers. Comput. Electron. Agric.; 2018; 144, pp. 102-113. [DOI: https://dx.doi.org/10.1016/j.compag.2017.11.032]

22. Zhu, W.; Peng, Y.; Ji, B. An Automatic Dead Chicken Detection Algorithm Based on SVM in Modern Chicken Farm. Proceedings of the 2nd International Symposium on Information Science and Engineering (ISISE 2009); Shanghai, China, 26–28 December 2009; IEEE Computer Society: Washington, DC, USA, 2009; pp. 323-326. [DOI: https://dx.doi.org/10.1109/ISISE.2009.62]

23. Pirompud, P.; Sivapirunthep, P.; Punyapornwithaya, V.; Chaosap, C. Application of Machine Learning Algorithms to Predict Dead on Arrival of Broiler Chickens Raised Without Antibiotic Program. Poult. Sci.; 2024; 103, 504. [DOI: https://dx.doi.org/10.1016/j.psj.2024.103504] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38335671]

24. Bao, Y.; Lu, H.; Zhao, Q.; Yang, Z.; Xu, W. Detection System of Dead and Sick Chickens in Large Scale Farms Based on Artificial Intelligence. Math. Biosci. Eng.; 2021; 18, pp. 6117-6135. [DOI: https://dx.doi.org/10.3934/mbe.2021306] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34517526]

25. Bochkovskiy, A.; Wang, C.Y.; Liao, H.Y.M. YOLOv4: Optimal Speed and Accuracy of Object Detection. arXiv; 2020; arXiv: 2004.10934v1

26. Wang, C.Y.; Bochkovskiy, A.; Liao, H.Y.M. YOLOv7: Trainable Bag-of-Freebies Sets New State-of-the-Art for Real-Time Object Detectors. arXiv; 2022; arXiv: 2207.02696v1

27. Liu, Y.; Kong, L.; Chen, G.; Xu, F.; Wang, Z. Light-Weight AI and IoT Collaboration for Surveillance Video Pre-Processing. J. Syst. Archit.; 2021; 114, 101934. [DOI: https://dx.doi.org/10.1016/j.sysarc.2020.101934]

28. Huang, Q.; Jiang, F.; Tong, Q.; Ma, Y.; Liu, Z.; Yang, X.; Wang, L. Knowledge Distillation Facilitates the Lightweight and Efficient Plant Diseases Detection Model. Plant Phenomics; 2023; 5, 62. [DOI: https://dx.doi.org/10.34133/plantphenomics.0062]

29. Cai, X.; Zhu, Y.; Liu, S.; Yu, Z.; Xu, Y. FastSegFormer: A Knowledge Distillation-Based Method for Real-Time Semantic Segmentation of Surface Defects in Navel Oranges. Comput. Electron. Agric.; 2024; 217, 108604. [DOI: https://dx.doi.org/10.1016/j.compag.2023.108604]

30. Belinga, A.G.; Tekouabou Koumetio, C.S.; El Haziti, M.; El Hassouni, M. YOLO-Based Masking and Deep CNN for People Counting Using Single Overhead Camera. J. Sens. Actuator Netw.; 2023; 12, 61. [DOI: https://dx.doi.org/10.3390/jsan12040061]