Content area
Jadeite jade, renowned for its unique texture and cultural significance, stands as the epitome of jade varieties, embodying the latest evolution of China's jade culture. This research endeavors to establish an AI model for precisely screening jadeite quality, employing deep learning techniques to revolutionize jadeite design and detection. The objective is to provide jewelry companies, designers, and customers with an unbiased means of grading and evaluating jadeite quality. We have meticulously curated a database of jadeite images, applied preprocessing techniques, and have harnessed convolutional neural networks (CNN) for feature extraction. The outcomes were promising, with the model achieving notable performance indicators: an accuracy rate of approximately 84.75%, a recall rate of about 84.94%, and an F1 score of roughly 73.76% in jade image classification tasks. These results underscore the model's effectiveness in the assessment of jadeite quality. Incorporating computer-aided technology into jadeite screening foreshadows a transformative era where artificial intelligence seamlessly integrates with traditional jade carving design, signifying a pivotal shift in the industry's landscape.
Jadeite jade, renowned for its unique texture and cultural significance, stands as the epitome of jade varieties, embodying the latest evolution of China's jade culture. This research endeavors to establish an AI model for precisely screening jadeite quality, employing deep learning techniques to revolutionize jadeite design and detection. The objective is to provide jewelry companies, designers, and customers with an unbiased means of grading and evaluating jadeite quality. We have meticulously curated a database of jadeite images, applied preprocessing techniques, and have harnessed convolutional neural networks (CNN) for feature extraction. The outcomes were promising, with the model achieving notable performance indicators: an accuracy rate of approximately 84.75%, a recall rate of about 84.94%, and an F1 score of roughly 73.76% in jade image classification tasks. These results underscore the model's effectiveness in the assessment of jadeite quality. Incorporating computer-aided technology into jadeite screening foreshadows a transformative era where artificial intelligence seamlessly integrates with traditional jade carving design, signifying a pivotal shift in the industry's landscape.
(ProQuest: ... denotes formulae omitted.)
INTRODUCTION
Initially, the term 'jadeite' denoted a vivid avian creature embellished with plumage in shades of red and green. The use of jade in various contexts is derived from the gem's resemblance to the vibrant appearance of the bird, characterized by its red, green, and purple hues.1 Jadeite jade is widely recognized as the most esteemed type of jade due to its exceptional qualities. It undergoes a process of metamorphic crystallization and subsequent recrystallization, originating from its progenitors, who are rich in sodium and aluminum. These transformations occur under high pressure and relatively mild temperatures.2 The gem's tactile qualities, characterized by its comfortable and warm texture, hold significant importance in Chinese aesthetics, which is shown in the substantial portion of its trade closely linked to both the Chinese market and the diasporic Chinese communities. According to data from the China Jewelry and Jade Jewelry Industry Association, the market is estimated to have a value of approximately US$101 billion in 2021. The constraints faced by the widespread standardization of jade items and the difficulties in building a comprehensive research and development framework for jade products. Our investigation, carried out within an evolving retail landscape, has revealed a pivotal change in the buying behavior of Chinese consumers, who increasingly value abstract principles over just utilitarian factors when making purchases.3 This trend points to a more profound transformation in the mindset of consumers in China, signaling a move towards a more aspirational and ideologically driven market. The advent of the digital information age presents a significant challenge to the cultural industry, posing a potential risk of cultural uniformity. However, it also catalyzes the pursuit of incorporating contemporary values and beliefs into the traditional jade culture. A lengthy and iterative process of learning and adaptation characterizes this endeavor.
The market size of China's jewelry and jadeite jewelry industry is approximately US$114,472 million. Jadeite is traded in huge volumes as a jewelry material in China, and the assessment of conventional jadeite quality relies on manual labor and conventional instruments. The main objectives of establishing an artificial intelligence (AI) model for jadeite quality screening include accurate screening to improve efficiency for the jadeite industry and using deep learning technology to accurately screen and assess the quality of jadeite, providing an unbiased grading and assessment method. This method increases the objectivity and fairness of the analysis and reduces human bias, and provides data-based decision support to jewelry companies, designers, and customers to help them better understand and evaluate the quality of jadeite. Developing new computer vision algorithms to extract unique features and quality indicators from jadeite photographs and to categorize these findings with neural network algorithms are the aims of this study.
Our research focuses on integrating deep learning and computer vision in jade image classification and quality prediction, marking the emergence of a cooperative partnership between algorithmic precision and human creativity. We are the first to combine jadeite image categorization and quality discrimination using deep machine learning techniques, using AI technology to expand its application to encompass the realm of arts and crafts. This extension enables a functional connection that elucidates the intricate effects of different elements on the quality and differentiation of jadeite. Furthermore, the algorithm employed integrates a selflearning mechanism alongside data augmentation techniques, facilitating a more profound comprehension and improved fidelity in representing authentic jadeite, and effectively tackling a prevalent challenge in computerized deep learning: the issue of data overfitting. We have adopted a comprehensive approach to mitigate this concern by engaging industry specialists, laboratory assessment professionals, and university professors. These individuals were involved in a rigorous process of methodically validating the raw data. This methodology mitigates the model's risk of acquiring an overly narrow focus through its exposure to the training data.
The primary findings of this study can be succinctly outlined as follows:
1. A thorough examination of jadeite imagery, using several existing techniques, such as physical observation and microscopic inspection, which may need to be improved for comprehensive analysis. Using deep learning techniques to evaluate jadeite quality yields increased objectivity and impartiality.
2. The inclusion of many algorithms in this study serves not only to provide a computer-based linguistic description of various jadeite properties but also to improve mathematical accuracy in terms of prediction and assessment.
3. Incorporating a self-learning technique for data augmentation enhances the comprehension and depiction of jadeite quality, consequently bolstering the prediction capabilities of the model.
4. The mitigation of overfitting is a prevalent issue encountered in machine learning. The performance of the jadeite classification model can be enhanced on novel, unobserved data by incorporating a broad spectrum of jadeite properties within the algorithmic framework.
The combined contributions made in this study significantly strengthen our capacity to assess and forecast the quality of jadeite, hence holding considerable consequences for both the jewelry business and jade carving designers. The application of computer vision algorithms for extracting distinctive characteristics and quality indicators from photographs of gemstones, and the subsequent categorization of these findings using network algorithms, is a novel approach to the field of jade carving. The study is organized into distinct sections that explore the historical development of jadeite, the assessment of its market value, and the fundamental research inquiries. Subsequently, a comprehensive examination of the existing body of literature on the recognition and design of jadeite images is presented. The following parts provide an overview of the experimental design and include an analysis of the empirical research findings. Ultimately, this research elucidates the substantial impact of deep learning on the efficacy of conventional jadeite identification, paving a promising trajectory for the advancement of jadeite design in the future.
RELATED WORK
Oriental cultures worldwide admire jadeite in an unparalleled way. However, most of this information remains confined to Chinese culture, and so unavailable to most Western individuals. As a result, the elusive nature of accessing jadeite adds to its mystique, enhancing the gemstone's enigmatic reputation among aficionados and collectors.4
Traditional Industry Classification of Texture, "Type-Wateriness," and Translucency
Jadeite consists of various minerals, including jadeite itself. Any change in the composition of these minerals results in corresponding changes in species quality, translucency, and color. These modifications result in several classifications of jadeite and offer boundless opportunities for creativity in jadeite design and carving.
The quality of jadeite is determined by the collective impact of its density, structural refinement, and level of transparency, this last being influenced by its roughness.5 The term "quality" pertains to the thickness of jadeite's crystal particles and the density of its crystal structure, commonly referred to within the industry as the thickness of the jade flesh. The categorization of jadeite's quality is a complex process that aims to promote its exchange within the trade industry. Various methods have been developed to classify jadeite based on color, transparency, and particle size. Examples of different types of jadeites include glass-type jadeite, ice-type jadeite, glutinous rice soup-textured jadeite, and bean-grain-type jadeite. Glass-type jadeite is widely recognized for its outstanding quality, distinguished by its exceptional clarity and transparency, which is comparable to that of conventional glass, which exhibits a high degree of transparency. The degree of transparency jadeite exhibits is rare, considering its intrinsic polycrystalline structure. A photograph of polished glass-type jadeite under reflective lighting is shown in Fig. 1A. Ice-type jadeite is considered a distinct variety, with its quality ranked second to that of glass-type jadeite. When comparing glass-type jadeite with ice-type jadeite, it can be seen that the latter exhibits a tiny cloudiness like ice, hence preventing it from achieving perfect transparency. However, ice-type jadeite can still evoke ice and jade sensations in individuals. A photograph of polished ice-type jadeite under reflective lighting is shown in Fig. IB. The culinary preparation known as glutinous rice soup exhibits a distinctive texture reminiscent of jadeite. The texture of the substance exhibits both clarity and turbidity, reminiscent of the aqueous medium employed in the process of rinsing sticky rice. The value of this jadeite is comparatively less compared to jadeites of the glass and ice types. A photograph of polished glutinous rice soup-textured jadeite under reflective lighting is shown in Fig. 1C. The jadeite variety known as bean-grain exhibits a translucent to somewhat transparent appearance, characterized by the presence of short columnar crystals that resemble the shape and size of beans. The substance has a notable presence of particles of larger dimensions, resulting in a frequently diminished level of transparency. The bean-grain-type jadeite is the predominant variety, constituting a significant proportion of commercially available jadeite. A photograph of polished bean-grain-type jadeite under reflective lighting is shown in Fig. 1D.1,6
The concept of "type-wateriness," is a valuable perspective for understanding the inherent characteristics of jadeite and its significance as a cultural object and one of its most enigmatic facets. Essentially, it can be regarded as a cultural descriptor that delineates the degree and caliber of translucency and texture.7 The primary aspect to consider in this classification is that the more the resemblance to "water" in a jadeite specimen, the higher its classification and, hence, its value.8 The categorization of these general classes are indicated in Table I. It is imperative to underscore that, within the disciplines of gemology and mineralogy, no established formal characterization exists that incorporates saturation and water content for jade; however, traders opt to depict jade constructs and articulate conceptions of "water" or its associated attributes.
Jade and Computer Vision
The field of computer vision presents a novel methodology for categorizing and predicting gemstone quality. The utilization of image-based processing techniques has gained significant traction and has been widely applied in several sectors.9 Various industries, including printing, processing, manufacturing, the medical sector, the food try, and have been studied by researchers.10-12 In this study, computer analysis software was designed to identify gathered photographs of green jadeite.13 The program utilized color data obtained from color digital cameras as its basis for analysis. The software can discern the categorization and chromatic consistency of jadeite. In conclusion, we have successfully shown the method's capacity to assess the hue of green jadeite efficiently, impartially, and consistently through subjective classification and validation using Gem Dialogue cards.
Researchers have delved into the assessment of jadeite color through an in-depth study of its physical properties, basic theoretical concepts, and implications for real-world applications.14 The findings proposed that the green grading system for jadeite should be grounded on the CIE 1976 L·a·b· uniform color space,15 which is a color-opponent space that is widely used to describe all the colors visible to the human eye. It was developed by the International Commission on Illumination (CIE) to provide a more uniform color space for color measurement and comparison. This approach has been supported by several studies that have emphasized the utility of the CIE 1976 L·a·b· color space in accurately capturing the nuances of jadeite color. Lightness (L·) represents the lightness of the color, with a range from 0 to 100, where 0 is black and 100 is white. Of the other two designations, a· greenred axis) represents the position between green and red, and b· (blue-yellow axis) represents the position between blue and yellow.15-17 Laser ablation inductively coupled-plasma mass spectrometry has been leveraged to compile an extensive database cataloging the trace elements found in jadeite from a variety of sources, contributing valuable insights into the field of gemology,18 with detailed analysis further discussed in that study. Following this, we have utilized a fusion technique that included a weighted extreme learning machine, AdaBoost, and incremental learning to construct a model to determine the origin of jadeite. This approach has successfully enabled intelligent classification of jadeite sources.
Computer vision has been applied in several subdomains within Earth sciences. One example of a mineral identification and classification technique was online ore sorting, which involved the examination of color and structure.19 A study was conducted in color analysis wherein color data were collected from 500 marble surface pictures.20 The data included red, green, and blue codes and hue, saturation, and intensity values. The researcher conducted a study wherein 1200 photos were taken from 120 limestone samples. Five main component features were identified from these images. Artificial neural networks have been employed in previous research to classify ore grade features, hence yielding dependable forecasts of ore grade. In another study, support vector machines were utilized to classify 12 different varieties of granite based on extracted color and texture parameters.21 By evaluating seven optical properties of mineralscolor, polychrome, interference color, birefringence, opacity, isotropy, and extinction angle-they an extensive database of 45 standard thin sections has been compiled.22 This endeavor facilitated the creation of a computerized mineral identification system that relies on color as a determining factor.
In jadeite trading, the concept of 'qualia'-the individual experience of subjective qualities-may be considered a pivotal element influencing market management and the jadeite market's economic interplay.23 In modern China, jadeite functions as both a cultural symbol and a semi-precious commodity, and its value in the jewelry sector is closely tied to its perceived quality. 4 The price of jadeite in the jewelry business strongly correlates with its quality. A notable disparity in price exists between jadeite of superior quality and jadeite of inferior quality. Hence, it is imperative for both jewelers and consumers to accurately forecast the quality of jadeite to establish equitable pricing and to facilitate the standardization of the jadeite jewelry industry. Moreover, jade carving designers need to possess a profound comprehension of the quality of jadeite when crafting jewelry items. The color, transparency, and texture of jadeite are crucial factors in determining the design's ability to effectively highlight and enhance the distinctive attributes of the gemstone. This factor significantly impacts their decision-making process in design and creative orientation. Computer deep-learning techniques enable the objective identification and classification of jadeite quality. The jadeite photographs utilized in this analysis were obtained from reputable domestic jewelry companies, presenting a diverse range of jadeite items in terms of quality. Images of specific jadeite products shown in Chinese live-streamed shopping sessions were also included. This technology has the potential to enhance the precise evaluation of jadeite quality while also playing a significant role in promoting the standardization and sustainable growth of the jadeite market.
METHODOLOGY
Data Collection and Preprocessing
The flowchart of jadeite image classification is shown in Fig. 2.
Data Source and Scale
This section presents the sources and scales of the jadeite image data used.
(1) Jade image data mainly comes from three sources:
* Online jewelry stores. We conducted a systematic image collection of some well-known online jewelry stores. These stores include "Emerald House," "Emerald Place," etc. All images obtained from these stores are used for research purposes only, and. to ensure the legality of the data, appropriate licenses or authorizations have been obtained for any copyrighted images in the store.
* Provided by partners. We have reached a cooperation agreement with the "Oriental Jade Research Institute," and they have provided us with many high-quality jade images. These images are intended primarily for academic research, and the Oriental Jade Research Institute has expressly agreed to the terms of use of the data.
* Other sources. Some data come from publicly available image datasets and volunteers' collections, all of which have secured permission or legality for research use.
(2) Data size:
* Number of images: 4000 images in total.
* Image resolution: most images have a resolution of 1024 x 1024 pixels.
* Image format: the original format is JPEG.
Variety and quality. The dataset and the jade images we collected include the four different jadeite varieties, which are divided into three categories according to their appearance and texture: glass-type jadeite, ice-type jadeite, glutinous rice soup-textured jadeite, and bean-grain-type jadeite. This diversity of data poses significant challenges to model training. Different color and texture characteristics mean that the model must accurately distinguish and identify complex jade properties. At the same time, this data diversity also ensures the model's broadness and robustness in practical applications.
Data Quality and Accuracy
(1) Data collection process
The data collection process combines automated and manual methods. Images of online jewelry stores were collected using automated crawler technology. The collection process was primarily through direct data exchange for data with partners and other institutions. We also organized a team of professional jewelry appraisers and photographers responsible for photographing jadeite on site to ensure the quality and authenticity of the images. Possible problems during the acquisition process included image reflections, shadows, background interference, and low-resolution issues in some online images.
(2) Data quality control
* Data deduplication. To ensure that images in the dataset were unique, we used feature hashing techniques to detect and remove duplicate images.
* Error repair or correction. Image processing tools repaired or corrected images disturbed by light, shadow, or background. Additionally, for low-resolution or poor-quality images, we removed them to ensure the overall quality of the dataset.
* Image enhancement. For some key jade images, we also applied image enhancement techniques, such as contrast adjustment, sharpening, etc., to improve their visualization quality.
(3) Data annotation
Data annotation is a critical step because it directly affects the quality of model training. Our annotation process was as follows:
* Annotation team. We assembled an annotation team that included jade identification experts and data scientists. They provided accurate species, color, and quality grade labels for each image.
* Annotation tool. We used a customized image annotation tool that supported multi-level label classification and provided image enlargement, comparison, and annotation functions to help annotators perform accurate annotation.
* Quality control. To ensure annotation accuracy, each image was independently annotated by at least two experts. If there were inconsistencies in their annotations, the image was submitted to a third jade identification expert for final judgment.
* The experience level of the annotators. Our annotation team members all had more than five years of experience in jade identification and have received special data annotation training. They have extensive knowledge of jade and a particular understanding of the basic concepts of data science and machine learning, ensuring the annotation process's efficiency and accuracy.
* Feature extraction using CNN. In this study, a CNN was employed for feature extraction from the annotated jadeite images, which are particularly effective in analyzing image data due to their ability to automatically learn hierarchical features, such as edges, textures, and patterns, directly from the input images. The CNN used in this research consisted of multiple convolutional layers, where each layer extracted increasingly complex features from the raw image data. These features were then used to build a model that could distinguish between different jadeite quality grades, species, and color variations. The CNN's ability to capture fine details and subtle differences in the images contributed significantly to the overall accuracy of the jadeite quality screening model.
The data classification is shown in Fig. 3.
Data Preprocessing Method
(1) Image size normalization
To ensure consistency of model inputs and reduce computational burden, we resized all images to the same resolution, specifically 256 x 256 pixels. To maintain image quality, we used bilinear interpolation for image resampling.
(2) Denoising process
Since jade images may be affected by the shooting environment and equipment, there are varying degrees of noise. To improve image quality, we adopted the following denoising methods:
* Median filter. The median filter is a nonlinear digital filtering technique that removes image noise, especially "salt and pepper noise". Given a pixel neighborhood P, the set of pixels is {p1,p2,...pn}, and the sorted set is {p(1),p(2),...,p(n)}. Then, the output pixel value after median filtering is the middle value, that is:
Pmedian = P(|n/2|) (1)
* Gaussian filter. For general random noise, the Gaussian filter can effectively reduce the noise while maintaining image details:
... (2)
* ·G(x,y) is the value of the two-dimensional Gaussian function, x and y are the lateral and longitudinal distance from the center of the Gaussian function, respectively, and a is the Gaussian function's standard deviation and determines the function's width or spread.
(3) Brightness adjustment
Since images may have been taken under different lighting conditions, we adjusted for brightness, contrast, and color balance. Specifically:
* Histogram equalization Improve the contrast of an image by adjusting its histogram.
* Brightness and contrast adjustment. We used linear transformation technology to make fine adjustments based on the overall brightness and contrast of the image.
* Color balance. We performed color correction for images with significant color casts to ensure that the true color of jadeite was accurately reflected.
(4) Missing data and outlier handling
* Missing data. Upon inspection of the dataset, we discovered that some images were corrupted due to issues during download or transfer. We chose to delete these missing data directly.
* Outliers. During the manual inspection, we found that some images were not emeralds but other gemstones or content unrelated to the project. These images were considered outliers and were removed after confirmation.
Deep Neural Network Architecture
Network Structure Design
The research chose a Conv Net-based architecture, which proved effective on multiple visual recognition tasks. This structure was chosen because the convolutional layer can extract local features from the image, while the pooling layer can reduce the spatial dimension and computational effort. The fully connected layer plays a decision-making role here. Given jadeite's diversity and subtle texture differences, we employed a deep structure to capture these details.
In this research, the following models were used:
(1) ResNet50:25
Concept: ResNet50 and other ResNet models are based on residual learning. Instead of learning unreferenced functions directly, these networks learn residual functions concerning the layer inputs.
Architecture: ResNetSO is a deep residual network with 48 Convolutional (Conv) layers and 1 MaxPool and 1 Average Pool layer.
Layers: It introduces shortcut connections that skip one or more layers without adding extra parameters or computational complexity.
(2) VGG:26
Concept: VGG models are known for their simplicity. They use only 3x3 Conv filters with stride and pad of 1 and 2x2 Max Pooling filters with stride.
Architecture: The VGG model comes in two main variants, VGG-16 and VGG-19, with 16 and 19 Conv layers, respectively.
Layers: VGG-16 includes 13 Conv layers and 3 fully connected layers.
(3) MobileV2:27
Concept: The network uses lightweight depthwise separable convolutions to filter features in the intermediate expansion layer
Architecture: MobileNetV2 has 53 Conv layers and uses an inverted residual structure where the input and output of the residual block are thin bottleneck layers opposite to traditional residual models.
Layers: It includes 1x1 Convolution with ReLU6, 3x3 Depth wise Convolution, and another 1x1 Convolution without any linearity for each block.
(4) Shuffle Net:28
Concept: The architecture optimizes the speedaccuracy tradeoff by reducing direct metrics, such as memory access costs, and increasing platformspecific performance rather than just minimizing FLOPs.
Architecture: ShuffleNet V2 is an efficient CNN designed for mobile devices that improves the speed and accuracy of the original ShuffleNet.
Layers: The provided source does not detail the layer structure of ShuffleNet V2, but it emphasizes the use of group convolution and channel shuffle operations to maintain high efficiency.
Loss Function and Optimization Algorithm
Loss function. Considering this is a multi-class classification task, we chose cross-entropy loss, which can measure the difference between the model's predicted probability distribution and the probability distribution of the true label, providing us with a clear optimization goal.
Optimization algorithm: We chose the Adam optimizer. Compared with traditional stochastic gradient descent (SGD), Adam combines the advantages of Adagrad and RMSprop. It can automatically adjust the learning rate, which usually converges faster and does not require manual adjustment of the learning rate.
Data Enhancement Strategy
Self-supervised learning has been a popular training strategy in recent years, in which the model is trained using unlabeled data and captures the rich structure in the data by solving some designed prediction tasks. In image processing, data augmentation methods can be considered a selfsupervised learning method that augments training data by transforming the original image and asking the model to predict the original image or some of its properties.
* Random rotation. Images of jade captured from different angles may show different details and features. We randomly rotated the image by an angle (for example, between - 15 and 15°) and let the model learn to recognize the relationship between the rotated image and the original image.
* Random cropping and scaling. To make the model more robust and able to identify the characteristics of jadeite from different sizes and regions, we performed random cropping and scaling operations.
* Random Flip. We flipped images horizontally or vertically to provide more variation in perspective and to increase the diversity of the data.
* Color Distortion. The color of jade may vary depending on light and environmental conditions. We slightly adjusted the image's brightness, contrast, and saturation so that the model could accurately identify jade under various lighting conditions.
* Affine transformation. To simulate the possible distortion of jade, we used affine transformation to slightly distort the image.
* Noise injection. Real-world images can be affected by varying degrees of noise. Adding slight random noise to the image could make the model more robust.
Through these data augmentation strategies, our model can learn the characteristics of jade from multiple angles, scales, and lighting conditions, thereby better generalizing to different jade images. This self-supervised data augmentation method increases the model's training data, improves the model's robustness, and helps prevent overfitting.
Training and Evaluation
Model Training Process
We optimized the model's weights during training to minimize the cross-entropy loss. The following is the specific training process:
* Selection of hyperparameters. Hyperparameters are parameters determined before training begins. We selected the best hyperparameters based on the performance of the validation set, including learning rate, weight decay, and dropout rate.
* Number of training iterations. The model was trained for multiple iterations, with each iteration traversing the entire training dataset. We selected enough iterations to ensure model convergence.
* Batch size. Batch size affects the speed and stability of model training. We chose an appropriate batch size to trade-off training speed and hardware resource utilization.
* Optimize model weights. We used the Adam optimizer to adjust the model weights to gradually reduce the training data's prediction error.
Evaluation Indicators
Key Metrics to Evaluate Model Performance Include
* Accuracy. This represents the ratio between the number of samples correctly predicted by the model and the total number of samples. This metric provides an overall measure of how well the model is performing across all classes. However, in cases of class imbalance, accuracy alone may not be sufficient to reflect the model's true effectiveness, as it could be biased towards the majority class. Accuracy gives an overall measure of how well the model is performing across all classes, but it might not be sufficient on its own in cases of class imbalance:
... (3)
* Recall. Also known as the true positive rate, recall measures the proportion of actual positive samples that are correctly identified by the model. A high recall indicates that the model is effective in identifying most of the positive samples, which is crucial in applications where missing positive cases (e.g., high-quality jadeite) is costly. Recall is critical for ensuring that the model captures as many positive instances as possible, reducing the risk of missing important positive cases:
... (4)
* Precision. Precision indicates the proportion of samples predicted to be positive by the model that are truly positive. This metric is particularly important in scenarios where the cost of false positives (e.g., misclassifying low-quality jadeite as high-quality) is high. High precision ensures that the model's positive predictions are reliable. Precision ensures that, when the model predicts a positive case, it is likely correct, minimizing the cost of false positives:
... (5)
* F1 Score. The Fl score is the harmonic mean of precision and recall, providing a balanced measure that considers both false positives and false negatives. The Fl score is especially useful when there is an uneven class distribution, as it combines the strengths of both precision and recall giving a more comprehensive assessment of the model's effectiveness. The Fl score balances precision and recall, offering a single metric that accounts for both false positives and false negatives, making it a valuable indicator when evaluating model performance in scenarios with class imbalance or varying costs of errors:
... 6
Cross-Validation Method
To ensure that the model performance evaluation is reliable, we used К-fold cross-validation:
* К-fold cross-validation. The entire dataset was randomly divided into К subsets. The training and validation process was repeated K times, using a different subset as the validation set and the remaining K-1 subsets as the training set. The final evaluation metric of the model is the average of K iterations.
* Prevents overfitting. K-fold cross-validation helps to prevent overfitting because the model no longer relies on a specific training and validation data split. Furthermore, by observing the performance difference for K iterations, we can estimate the stability and generalization ability of the model.
* Obtain reliable evaluation results. The results are more robust and trustworthy because the model was evaluated on multiple training-validation data splits.
In summary, these three parts ensured that the model weights were optimized, appropriate evaluation metrics were selected, and the K-fold crossvalidation method was used to obtain reliable evaluation results.
RESULTS AND DISCUSSION
Model Performance Results
Training Set and Validation Set Performance
* The image depicted in Fig. 4 is of the model's training and validation losses.
* High initial loss. As can be seen from the figure, initially, both training and validation losses were very high. This is normal because the model has just started training, and the parameters have not yet been optimized.
* Rapid decline. The loss decreased rapidly as training progressed, which meant that the model is learning and adapting to the data.
* Fluctuations in validation loss. While training loss was relatively stable and continued to decrease, validation loss increased during specific periods. This may mean that the model has overfitted the training data at some point, causing performance degradation on the validation data.
* Stable phase. After about 40 epochs, both loss curves stabilized without significant decline, which may have meant that the model had achieved optimal performance with the current data and structure.
* Loss gap. There was a particular gap between training loss and validation loss, but it was not very large, which meant that the model may be slightly overfitted, but not severely.
* Small fluctuations in later stages. Although the loss had become relatively stable in the later stages of training, small fluctuations could still be observed, which could be due to the learning rate, the randomness of the data within a batch, or other factors.
* Figure 5 shows four different images, each labeled with the model's prediction (Pred) and the true class label (True). For example, in the first image, Pred and True are the glass species, and the prediction for this image is correct. As can be seen, this may be a glass object with glass characteristics and texture.
Test Set Performance
The results of an independent test set evaluation for the simulated model's performance are shown in Table II.
Comparison with Other Methods
To show the performance comparison between the current model and other methods on the same test set, a comparison is shown in Table III.
When comparing the network structures based on the given performance metrics, ResNet50 shows a clear advantage in all categories: accuracy, recall, precision, and Fl score. ResNet50 is more adept at generalizing from the training data and correctly classifying jade images. VGG's performance could be much better, possibly due to its simpler architecture not capturing the complex features needed for accurate jade classification. MobileNetV2 and Shuffle Net offer moderate performance, with MobileNetV2 slightly outperforming Shuffle Net. Both are designed for efficiency and may trade off some accuracy for speed and lower computational resource usage.
The high F1 score for MobileNetV2 relative to its accuracy suggests a balanced performance between precision and recall. ResNet50 is not as good, with the Fl score being notably lower than the accuracy and recall, which indicates a discrepancy between precision and recall in ResNet50's predictions. Shuffle Net's lower scores across all metrics compared to ResNet50 and MobileNetV2 might be due to its focus on reducing computational costs, which may come at the cost of model complexity and the feature extraction capability needed for this task. Impact of Data Augmentation
Data augmentation is a technique that augments a training set by creating a modified version of the original data. It can effectively increase the diversity of the training data, thereby improving the model's generalization ability. The following is a detailed description and discussion of the impact of data augmentation on model performance.
To improve the model's generalization ability, we applied the following data augmentation strategy to the original dataset:
Random rotation ± 20°
Horizontal flip
Randomly cut 10%.
Brightness adjustment ± 10%
After applying data augmentation, our dataset was expanded to 3000 images, and, after training with the augmented dataset, the model's accuracy on the validation set increased to 89%. Data augmentation is an effective means of improving the model's performance and generalization ability. However, it is key to choosing an appropriate data augmentation strategy and verifying it with experiments.
Cross-Validation Results
Cross-validation is a technique for evaluating model performance, especially when the amount of data is limited. In K-fold cross-validation, the dataset was randomly divided into K subsets; each time, the K-1 subsets were used as training data, and the remaining 1 subset was used as validation data. This process was repeated K times, with each subset used once as validation data. Finally, we obtained K different performance evaluation results to calculate the mean and variance of the model performance.
Experimental Results
The accuracy obtained for each validation is set out in Table IV.
From these results, we can calculate that the mean of the average accuracy was 84.8%, while the variance indicates the degree to which accuracy fluctuated during these 5 verifications. A lower variance means that the model's performance was more stable.
The Impact of Cross-Validation on the Model
* Model stability. Variance can help us understand the stability of the model. In the above example, the accuracy variance was slight, which means that different data partitions bring little performance change to the model, so the model was relatively stable.
* Model reliability. The mean or average accuracy tells us the average performance of the model. In the above example, the average accuracy was 85%, which is relatively high, indicating that the model was more reliable.
Discussion
In this section, we will provide an in-depth analysis of the model's performance and consider some factors that may have affected it. At the same time, we will discuss the difficulty of different categories of jade images in model classification and consider possible application problems and limitations of the model.
(1) Model performance analysis
The model's performance was good, and the specific performance indicators were: the accuracy rate was about 84.75%, the recall rate was about 84.94%, and the Fl score was about 73.76%. These show that the model achieved a good overall jade image classification task performance.
(2) Factors that may affect model performance
* Data quality. Data quality can be one of the critical factors affecting model performance. The model may be affected by label errors, noise, or imbalance in the dataset. In practical applications, the dataset should be carefully cleaned and preprocessed.
* Data augmentation strategies. We tried different data augmentation strategies, including random cropping, horizontal flipping, etc. These strategies impact the model's performance, and different strategies may be suitable for different types of datasets and tasks.
* Model architecture. We used the ResNet-50 model, but other pre-trained models or custom architectures could be tried to improve performance. Model depth and complexity may also affect performance.
* Learning rate and optimizer. The choice of learning rate and optimizer is crucial for model training. Tuning the learning rate and optimizer hyperparameters can significantly impact performance.
(3) Ease of classification of different categories
* Easily classified categories. Some jade varieties may have unique characteristics that make them easily classified by the model. For example, jadeite of a specific color or texture may be easier to distinguish.
* Difficult to classify. Other jadeite varieties may have similar appearance characteristics, making them difficult to distinguish. This may require more data and more complex models to improve classification accuracy.
(4) Possible problems in practical applications and limitations of the model
* The difficulty of data collection. Collecting largescale, high-quality jade image data can be complicated because it requires specialized knowledge and equipment.
* Labeling issues. Determining the accurate label of a jadeite variety can be challenging because experts may have different opinions about the same piece of jadeite.
* Model generalization. A trained model may perform well on a specific dataset but may be limited in generalizing jade images from different sources or conditions.
* Computing resources. Deep learning models require significant computing resources, including GPUs. In some cases, computing resources may be limited.
Considering these factors, the model's performance and application must be carefully evaluated, and further improvement and optimization may be required to meet actual needs.
PROSPECTS
In the future, this research, the emerald industry, and technical computing will bring about many critical shifts. On the one hand, deep learning and computer vision have been integrated into jadeite quality assessment, and this integration can improve the objectivity and accuracy of jadeite grading. Developing AI models for grading jadeite quality could lead to greater standardization and consistency in quality assessment across the industry. This is important for building trust and reliability in the marketplace, as differences in quality assessments can lead to differences in pricing and value. Improved quality assessment could increase market operation efficiency and potentially influence industry trade dynamics.
On the other hand, utilizing a database of selected emerald images and deep learning techniques for feature extraction and quality assessment, could drive the industry towards data-driven decisionmaking. This could provide a more scientific and systematic approach to jadeite appraisal rather than relying solely on the expertise of individual appraisers. AI models can provide unbiased ratings and assessments, thereby reducing the likelihood of subjectivity in quality assessment. This is particularly important in an industry where jadeite's cultural significance and aesthetic value can affect valuation. This research has focused on using AI for jadeite design and detection, which could lead to innovative approaches to jadeite product creation and identification, which, in turn, could lead to new designs that more effectively utilize the unique attributes of jadeite, as well as more accurate and efficient detection methods.
AI combined with emerald screening can apply technology to traditional processes. In terms of traditional jade carving products, this technology for recognizing and assessing the quality of jadeite provides support for traditional artisans to create more works, and the artisans will not have rights issues with consumers over the quality of the jadeite. By making quality assessment transparent and standardized, AI helps to expand the market scope of traditional craftsmanship and make it more accessible to a global audience, thereby facilitating cultural exchange and cross-cultural appreciation, and, in terms of gentrification of the appreciation of the craft, unravelling the mystery of Chinese jadeite, reducing the cost of jadeite appraisal, and allowing more people to participate in and appreciate the art form of jadeite jade carving. Technology thus helps traditional crafts to adapt to modern lifestyles, bringing them closer to and integrating them into contemporary environments, ensuring their continued relevance and development, helping to make traditional crafts more economically viable, and supporting the livelihoods of artisans and the continuation of these practices.
CONCLUSIONS
Incorporating deep learning techniques for identifying and categorizing jadeite imagery marks a substantial advancement in jewelry design. The study's findings underscore the model's robust classification capabilities, demonstrated by impressive accuracy, precision, and Fl scores. They highlight its potential utility in enhancing jadeite visualization, evaluation, and the overall design process.
However, the research also illuminates critical factors influencing model performance. Data quality and preparation, selection of data augmentation techniques, architectural choices in model construction, and optimization parameters are instrumental in determining the efficacy of deep learning applications. Specific jadeite varieties present classification challenges due to shared characteristics, indicating a need for richer datasets and more intricate models to improve discernment accuracy. Practical application considerations, such as the complexity of data collection in this domain, subjective annotation affected by expert variance, and the ability of models to generalize across different datasets, have also been revealed in this study. These factors are important concerns that may affect the widespread application of deep learning techniques in the emerald jewelry industry.
REFERENCES
1. T. Hongyun. Quality and Standardization (12), 30-33 (2020).
2. Y. Guo, X. Zong, M. Qi, Y. Zhang, and H. Wang, EURASIP J. Image Video Process, https://doi.org/10.1186/s13640-0180342-2 (2018).
3. M.-J. Kim and Y.-J. Lee, Korean J. Hum. Ecol. 24(3), 423436 (2015).
4. R.W. Hughes, O. Galibert, G. Bosshart, F. Ward, T. Oo, M. Smith, T.T. Sun, and G.E. Harlow, Gems Gemol. 36(1), 2-26 (2000).
5. H. González-Jorge, V. Alvarez-Valado, J.L. Valencia, and S. Torres, Sensors 10(4), 4002-4009 (2010).
6. Y. Guo, China Gold Jewelry, 2011(5), 110-117 (2011).
7. H. Møller, Spectral Jade: Materiality, Conceptualisation, and Value in the Myanmar-China Jadeite Trade. Doctoral Thesis (monograph) (2019).
8. Y. Ma, Qualia in Markets: Ruili's Jadeite Marketplaces in the China-Myanmar Borderland University of Hawai'i at Manoa] (2023).
9. H. Tang, H. Wang, L. Wang, C. Cao, Y. Nie, and S. Liu, Jom 75(7), 2590-2602 (2023).
10. M. Häfner, M. Liedlgruber, A. Uhl, A. Vécsei, and F. Wrba, Med. Image Anal. 16(1), 75-86 (2012).
11. P. Vithu and J. Moses, Trends Food Sci. Technol. 56, 13-20 (2016).
12. S. Zhang and Y. Guo, Minerals, https://doi.org/10.3390/mi П11080791 (2021).
13. Z. Lili and Y. Xinqiang, J. Gems Gemmol. 19(3), 45-51 (2021).
14. Y. Guo, Bulg. Chem. Commun. 49(4), 961-968 (2017).
15. Y. Guo, X. Zong, and M. Qi, Multimed. Tools Appl. 78, 841856 (2019).
16. Y. Guo, H. Wang, and H. Du, Multimed. Tools Appl. 15, 14491-14502 (2016).
17. X. Pan, Y. Guo, Z. Liu, Z. Zhang, and Y. Shi, Earth Sei. Res. J. 23(4), 371-378 (2019).
18. L. Wang, J. Tu, Y. Li, M. Li, Artificial Intelligence, and Data Processing (ICCAID 2022) (2023).
19. P. Hofmann, R. Marschallinger, M. Unterwurzacher, and F. Zobl, Earth Sci. 106(2), 73-82 (2013).
20. M. Gökay and I. Gundogdu, Constr. Build. Mater. 22(7), 1342-1349 (2008).
21. F. Bianconi, E. Gonzalez, A. Fernandez, and S.A. Saetta, Expert Syst. Appl. 39(12), 11212-11218 (2012).
22. S. Aligholi, R. Khajavi, and M. Razmara, Comput. Geosci. 85, 175-183 (2015).
23. M. Callon, Social. Reo. 46(1), 1-57 (1998).
24. R.N. Proctor, Configurations 9(3), 381-412 (2001).
25. B. Koonce and B.E. Koonce, Convolutional neural networks with swift for tensorflow: Image recognition and dataset categorization (Springer, Berlin, 2021).
26. A. Vedaldi and A. Zisserman, Vgg convolutional neural networks practical (Department of Engineering Science, University of Oxford, Oxford, 2016), p66.
27. K. Dong, C. Zhou, Y. Ruan, Y. Li, MobileNetV2 model for image classification, in 2020 2nd International Conference on Information Technology and Computer Application (ITCA) (2020).
28. S. Ghosh, M. J. Mondal, S. Sen, S. Chatterjee, N. K. Roy, S. Patnaik, A novel approach to detect and classify fruits using ShuffleNet V2, in 2020 IEEE applied signal processing conference (ASPCON) (2020).
Copyright Springer Nature B.V. Jan 2025