A low-resolution industrial image classification method and device based on label-guided diversity contrast learning
By constructing a feature extractor-classifier framework and a label-guided diversity contrastive learning method, the problem of low accuracy in low-resolution image classification is solved, achieving efficient and accurate low-resolution image classification, which is suitable for rapid defect screening and real-time monitoring in industrial production.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUAZHONG UNIV OF SCI & TECH
- Filing Date
- 2025-04-14
- Publication Date
- 2026-07-07
AI Technical Summary
Most existing image classification methods are designed for high-resolution images and cannot effectively address the problem of decreased classification accuracy caused by low-resolution images.
We employ a label-guided diversity contrastive learning approach to construct a feature extractor-classifier framework. We use sequence-decoupled large-kernel convolutional modules and a label-guided diversity contrastive learning strategy to train a neural network model, achieving efficient and accurate classification of low-resolution images.
It improves the classification accuracy of low-resolution images, reduces the computational cost of data processing, and is suitable for rapid defect screening and real-time monitoring in large-scale production, while reducing hardware and computational costs.
Smart Images
Figure CN120451626B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image classification technology, and more specifically, relates to a method and device for classifying low-resolution industrial images based on label-guided diversity contrastive learning. Background Technology
[0002] With the rapid development of industrial automation and intelligent manufacturing, industrial image acquisition and analysis technologies have been widely used in production monitoring, quality inspection, and equipment maintenance. However, in actual industrial production processes, the resolution of acquired images is usually low due to limitations imposed by equipment, environment, and operating conditions. This low resolution leads to loss of image details and is often accompanied by noise and blurring, thus affecting the analysis and processing of downstream tasks.
[0003] In industrial visual inspection, the occurrence of low-resolution images is common and unavoidable. For example, in production line monitoring, due to the need to capture a large number of images in real time, the image resolution is usually compressed to a low level due to limitations in bandwidth, equipment processing power, and real-time requirements. In extreme environments such as high temperature, high pressure, and radiation, to avoid damage to equipment or operators, images are usually acquired from a distance or using sensors made of special materials, resulting in low image resolution and blurred details. Furthermore, in large-scale production, to reduce hardware costs and improve data processing efficiency, large-area images captured by mid-to-low-end cameras are often used for rapid pre-screening.
[0004] Image classification is a crucial task in industrial vision inspection, aiming to assist automated systems in making decisions through the effective classification and recognition of images. Image classification technology can solve many practical problems in industrial production. For example, in the manufacturing process, it can monitor the production line status in real time, promptly identify problems and issue corresponding alarms, reminding engineers to adjust processes, thereby improving product consistency and quality. In the outgoing quality inspection stage, it can distinguish between images of different categories of products, automatically separating qualified and unqualified products, reducing manual intervention and improving production efficiency. In equipment maintenance, it can detect minor damage or wear on equipment surfaces, enabling early warning and preventing major malfunctions and equipment downtime.
[0005] Low-resolution image classification plays an indispensable role in improving production efficiency, reducing costs, and enhancing product quality. Because low-resolution images have lower requirements for storage, transmission, and processing, using them for classification and detection can effectively reduce reliance on high-end image acquisition equipment, lowering system construction and operating costs. Furthermore, low-resolution image classification technology can significantly improve image processing efficiency, making it particularly suitable for rapid defect screening and real-time monitoring in large-scale production. Therefore, low-resolution image classification technology has significant strategic importance for promoting the intelligentization and automation of industrial production.
[0006] In conclusion, low-resolution image classification technology has significant engineering implications in the field of industrial automation. However, most existing image classification methods are designed for high-resolution images and cannot effectively address the challenges posed by low-resolution images, resulting in a significant decrease in classification accuracy in low-resolution scenes. Summary of the Invention
[0007] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a low-resolution industrial image classification method and device based on label-guided diversity contrastive learning, which aims to solve the problem of low classification accuracy of existing low-resolution images.
[0008] To achieve the above objectives, according to one aspect of the present invention, a low-resolution industrial image classification method based on label-guided diversity contrastive learning is provided, the classification method comprising the following steps:
[0009] (1) The feature extractor in the neural network model is trained using pre-acquired low-resolution industrial image data with category labels and a diversity contrast learning strategy. The trained feature extractor extracts the representation vectors of low-resolution images of known categories, and the extracted representation vectors are used to train the classifier in the neural network model to obtain a low-resolution image classification model. The feature extractor is embedded with a sequence decoupling large kernel convolution module.
[0010] (2) Input the low-resolution image to be classified into the obtained low-resolution image classification model. The feature extractor extracts features from the corresponding low-resolution image and generates a representation vector. The classifier decodes the representation vector and outputs the classification result to realize the automatic classification of low-resolution images.
[0011] Furthermore, the feature extractor uses a convolutional neural network structure, and the classifier uses a feedforward neural network structure, which is used to decode the representation vector from the feature extractor layer by layer, thereby achieving the classification of low-resolution images.
[0012] Furthermore, the feature extractor uses ResNet18 as the backbone, and the classifier consists of three layers: an input layer, a hidden layer, and an output layer.
[0013] Furthermore, the sequence decoupling large kernel convolution module comprises two parts: sequence decoupling and sequence space selection. Sequence decoupling uses depth-wise convolution sequences. Explicitly decouple the large kernel convolution mechanism, and then use Ghost convolution. The channel suppression of each depth-wise convolution process is restored; the sequence space selection first concatenates the feature maps H extracted from all receptive fields along the channel dimension. i Furthermore, global average pooling and global max pooling are performed on the overall feature map obtained by concatenating the feature maps extracted from receptive fields of all scales along the channel dimension through parallel channels. This yields two pixel-level pooled feature matrices; then, depooling convolution is used. The number of channels is restored to the sequence value N, and the sigmoid function is used to activate the feature map after depooling convolution, obtaining the mask weights for each large kernel decomposition sequence; finally, depooling convolution is used. After restoring the number of input channels, the input X is then connected to the residual. i Perform the Hadamard product to obtain the final output value Y. i .
[0014] Furthermore, the final output value Y i The formula is:
[0015]
[0016] The classifier includes a dropout layer to apply regularization constraints.
[0017] Furthermore, the input to the feature extractor is the original low-resolution image x. i and its corresponding category label y i The output is a representation vector containing discriminative information. The parameters θ of the neural network are optimized by minimizing the distance between the representation vectors of similar images and maximizing the distance between the representation vectors of dissimilar images.
[0018] Furthermore, the label-guided diversity contrastive learning strategy is divided into two parts: label-guided contrastive learning and diversity contrastive learning. Label-guided contrastive learning learns a low-dimensional representation by maximizing the consistency of positive sample pairs and minimizing the consistency of negative sample pairs. Given a dataset There are a total of n samples The loss function for label-guided contrastive learning is:
[0019]
[0020] in, and Let x represent the positive sample pair set and the negative sample pair set, respectively. P It is a positive sample pair of sample x, x N Let f(·θ) be the negative sample pair of sample x, and f(·θ) be the mapping relationship learned by the neural network.
[0021] Furthermore, diversity contrastive learning introduces an additional quadratic regularization term based on the Gram matrix into the loss function of label-guided contrastive learning. To penalize similarity and promote orthogonalization of the weight matrix W to learn more diverse representations, the loss function for label-guided diversity contrastive learning is:
[0022]
[0023] The present invention also provides a low-resolution industrial image classification system based on label-guided diversity contrastive learning. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the low-resolution industrial image classification method based on label-guided diversity contrastive learning as described above.
[0024] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the label-guided diversity contrastive learning-based low-resolution industrial image classification method as described above.
[0025] In summary, compared with the prior art, the low-resolution industrial image classification method and device based on label-guided diversity contrastive learning provided by this invention have the following beneficial effects:
[0026] 1. This invention constructs a neural network model LR-CLS with a feature extractor-classifier framework, and introduces a sequence-decoupled large kernel convolution module and a label-guided diversity contrastive learning strategy to achieve efficient and accurate classification of low-resolution industrial images. This overcomes the dependence of traditional image classification methods on high-resolution images, reduces the computational cost of data processing, and improves the accuracy of low-resolution image classification.
[0027] 2. This invention proposes a sequence-decoupled large kernel convolution module, which uses depth-wise convolution sequences to independently extract spatial features and restores channel suppression through the ghost module, thus explicitly decoupling the classic large kernel convolution. Then, through a sequence spatial selection mechanism, the optimal filter is dynamically selected based on the foreground size of the input image, achieving low-cost and adaptive large-scale feature interaction.
[0028] 3. This invention employs a label-guided diversity contrastive learning strategy, which utilizes prior knowledge to optimize the training process of classical self-supervised contrastive learning and promotes the orthogonalization of the weight matrix through a quadratic regularization term based on the Gram matrix to obtain diverse representations, thereby improving the model's ability to extract discriminative features under low-resolution conditions.
[0029] 4. Diversity-guided contrastive learning introduces an additional quadratic regularization term based on the Gram matrix into the loss function of label-guided contrastive learning. To penalize similarity, the weight matrix W is orthogonalized to learn more diverse representations, thereby improving the model's ability to extract features from low-resolution images.
[0030] 5. Sequence decoupling using depth-wise convolution sequences Explicitly decouple the large kernel convolution mechanism, and then use Ghost convolution. By restoring channel suppression in each depth-wise convolution process, the input feature map is realized. Low-cost, wide-ranging information exchange.
[0031] 6. The classifier includes a dropout layer to apply regularization constraints and avoid overfitting. Attached Figure Description
[0032] Figure 1 This is a flowchart of a low-resolution industrial image classification method based on label-guided diversity contrastive learning provided by the present invention;
[0033] Figure 2 This is a diagram of the LR-CLS network structure constructed according to an embodiment of the present invention;
[0034] Figure 3 This is a schematic diagram of the tag-guided diversity contrastive learning strategy in an embodiment of the present invention. Detailed Implementation
[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0036] This invention provides a low-resolution industrial image classification method based on label-guided diversity contrastive learning. The classification method constructs a neural network model with a feature extractor-classifier framework and introduces a sequence-decoupled large-kernel convolution module and a label-guided diversity contrastive learning strategy to achieve efficient and accurate classification of low-resolution industrial images.
[0037] This invention, based on deep learning technology, develops a low-resolution image classification model, achieving efficient automatic classification of low-resolution images. This method offers low cost and high efficiency in signal analysis, while also demonstrating excellent classification accuracy.
[0038] Please see Figure 1 The classification method mainly includes the following steps:
[0039] S1. Construct a deep learning-based neural network model, LR-CLS, which employs a feature extractor-classifier framework. The feature extractor uses a convolutional neural network (CNN) structure to extract effective features from low-resolution images and transforms these features into high-order representation vectors containing discriminative information. The classifier uses a feedforward neural network (FFN) structure to decode this representation vector layer by layer, thereby achieving accurate classification of low-resolution images.
[0040] like Figure 2 As shown, the feature extractor uses ResNet18 as the backbone and embeds a sequence-decoupled large kernel convolution module; the classifier uses an FFN structure, consisting of three layers: input layer, hidden layer, and output layer.
[0041] The feature extractor uses a CNN with residual connections as its backbone and embeds a sequence-decoupled large-kernel convolution module, enabling the feature extractor to adaptively and dynamically select the most suitable feature map for the size of the object to be detected from large kernels of different scales. Specifically, the sequence-decoupled large-kernel convolution module consists of two parts: sequence decoupling and sequence space selection.
[0042] Sequence decoupling uses depth-wise convolutional sequences Explicitly decouple the large kernel convolution mechanism, and then use Ghost convolution. By restoring channel suppression in each depth-wise convolution process, the input feature map is realized. Low-cost, wide-ranging information exchange:
[0043]
[0044] Sequence space selection first involves concatenating feature maps H extracted from receptive fields of all scales along the channel dimension. iFurthermore, global average pooling and global max pooling are performed on the overall feature map obtained by concatenating the feature maps extracted from receptive fields of all scales along the channel dimension through parallel channels. This yields two pixel-level pooled feature matrices. Then, depooling convolution is used. The number of channels is restored to the sequence value N, and the sigmoid function is used to activate the depooling convolution after restoration, obtaining the mask weights for each large kernel decomposition sequence. Finally, depooling convolution is used... After restoring the number of input channels, the input X is then connected to the residual. i Perform the Hadamard product to obtain the final output value Y. i :
[0045]
[0046] The classifier includes a dropout layer to apply regularization constraints and prevent overfitting. The dimension of the representation vector output by the feature extractor is positively correlated with the resolution of the input image.
[0047] S2 utilizes a pre-acquired low-resolution industrial image dataset with clearly defined category labels. The feature extractor is trained in the neural network model. Specifically, the input to the feature extractor is the original low-resolution image x. i and its corresponding category label y i The output is a high-order representation vector containing discriminative information. The training process employs a contrastive learning approach, which optimizes the neural network parameters θ by minimizing the distance between the representation vectors of similar images and maximizing the distance between the representation vectors of dissimilar images.
[0048] In this embodiment, the input is an image x with a resolution of 32×32. i and its category label y i The output is a 128-dimensional representation vector. Cosine similarity is used to quantify the distance between different image representation vectors. The system is trained for 1000 generations with an initial learning rate of 0.5, which decays by 50% at the 500th and 850th generations. A label-guided diversity contrastive learning strategy is used as the loss function, and the model parameters θ are optimized generationally using an SGD optimizer.
[0049] A diagram illustrating a label-guided diversity contrastive learning strategy is shown below. Figure 3 As shown, it is divided into two parts: label-guided contrastive learning and diversity-guided contrastive learning. Label-guided contrastive learning learns a low-dimensional representation by maximizing the consistency of positive sample pairs and minimizing the consistency of negative sample pairs. Assuming dataset There are a total of n samples The loss function for label-guided contrastive learning is:
[0050]
[0051] in, and Let x represent the positive sample pair set and the negative sample pair set, respectively. P It is a positive sample pair of sample x, x N Let f(·θ) be the negative sample pair of sample x, and f(·θ) be the mapping relationship learned by the neural network.
[0052] Diversity contrastive learning introduces an additional quadratic regularization term based on the Gram matrix into the loss function of label-guided contrastive learning. To penalize similarity, the weight matrix W is orthogonalized to learn more diverse representations, thereby improving the model's feature extraction ability for low-resolution images. Therefore, the loss function for label-guided diversity contrastive learning is:
[0053]
[0054] During the training of the feature extractor, a cosine similarity metric is selected to quantify the distance between different image representation vectors, and an SGD optimizer is selected to optimize the model parameters of the feature extractor generation by generation.
[0055] S3 involves extracting representation vectors of low-resolution images of known categories using a trained feature extractor, and then using these extracted representation vectors to train a classifier in the neural network. Specifically, the classifier's input is the representation vector of the low-resolution image. and its category label y i The output is the predicted category of the image. During training, a cross-entropy loss strategy is used to minimize the predicted class. With real label y i The differences between them are used to optimize the parameters θ of the neural network.
[0056] During the training process of the classifier, the Adam optimizer is selected to optimize the model parameters of the classifier generation by generation.
[0057] In this embodiment, the input is a 128-dimensional representation vector. and its category label y i The output is the predicted category of the corresponding image. The input layer has a dimension of 128, the hidden layer has a dimension of 64, and the output layer has a dimension equal to the number of classes. The first two layers have a dropout coefficient of 0.2, and the training is performed for 100 generations with a learning rate of 0.001. Cross-entropy loss is used as the loss function, and the model parameters θ are optimized generation by generation using the Adam optimizer. The definition of cross-entropy loss is:
[0058]
[0059] S4 calls the model parameters of the feature extractor trained in S2 in the feature extractor part of the neural network model framework built in S1, and calls the model parameters of the classifier trained in S3 in the classifier part, to obtain the complete low-resolution image classification model LR-CLS.
[0060] S5 inputs the low-resolution image to be classified into the low-resolution image classification model obtained in S4. The feature extractor extracts features from the low-resolution image and generates a representation vector. The classifier decodes the representation vector and outputs the classification result, thus realizing the automatic classification of low-resolution images.
[0061] During the classification process, the resolution of the image to be classified should be consistent with the resolution of the image used for model training.
[0062] The present invention will be further described in detail below with reference to specific embodiments.
[0063] To verify the practical application effect of the present invention, the present invention was validated on the Northeastern University Hot Rolled Steel Surface Defect Dataset (NEU-CLS) and the Chongqing University Weld Joint Surface Defect Dataset (RSW-C). Detailed information on the two datasets is summarized in Table 1.
[0064] Table 1 Dataset Sample Information
[0065]
[0066]
[0067] The state-of-the-art (SOTA) models MFN-ResNet50 from the original paper on the NEU-CLS dataset and FCA-ResNet50 from the original paper on the RSW-C dataset were selected as control groups to fully verify the effectiveness of the proposed method. During the experiments, the input image resolution was 32×32, and each experiment underwent five independent training runs. The average result of the five runs was used as the final evaluation metric. The experimental results are summarized in Table 2.
[0068] Table 2 Classification results of low-resolution images
[0069]
[0070] As clearly shown in Table 2, compared with other image classification methods, the LR-CLS model provided in this invention performs better in low-resolution image classification tasks, achieving accuracies exceeding those of state-of-the-art (SOTA) models by 1.33% and 8.46% on the NEU-CLS and RSW-C datasets, respectively. Therefore, the method proposed in this invention can effectively improve the classification accuracy of low-resolution images. Furthermore, the method proposed in this invention is applicable to more diverse industrial scenarios and can reduce the hardware costs of industrial image acquisition and the computational costs of signal analysis.
[0071] The present invention also provides a low-resolution industrial image classification system based on label-guided diversity contrastive learning. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the low-resolution industrial image classification method based on label-guided diversity contrastive learning as described above.
[0072] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the label-guided diversity contrastive learning-based low-resolution industrial image classification method as described above.
[0073] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A low-resolution industrial image classification method based on label-guided diversity contrastive learning, characterized in that, This classification method includes the following steps: (1) The feature extractor in the neural network model is trained using pre-acquired low-resolution industrial image data with category labels and a diversity contrast learning strategy. The trained feature extractor extracts the representation vectors of low-resolution images of known categories, and the extracted representation vectors are used to train the classifier in the neural network model to obtain a low-resolution image classification model. The feature extractor is embedded with a sequence decoupling large kernel convolution module. (2) Input the low-resolution image to be classified into the obtained low-resolution image classification model. The feature extractor extracts features from the corresponding low-resolution image and generates a representation vector. The classifier decodes the representation vector and outputs the classification result to realize the automatic classification of low-resolution images. The sequence decoupling large kernel convolution module consists of two parts: sequence decoupling and sequence space selection. Sequence decoupling uses depth-wise convolution sequences. Explicitly decouple the large kernel convolution mechanism, and then use Ghost convolution. The channel suppression of each depth-wise convolution process is restored; the sequence space selection first concatenates the feature maps extracted from all receptive fields along the channel dimension. Furthermore, global average pooling and global max pooling are performed on the overall feature map obtained by concatenating the feature maps extracted from receptive fields of all scales along the channel dimension through parallel channels. This yields two pixel-level pooled feature matrices; then, depooling convolution is used. The number of channels is restored to the sequence value N, and the sigmoid function is used to activate the feature map after depooling convolution, obtaining the mask weights for each large kernel decomposition sequence; finally, depooling convolution is used. After restoring the number of input channels, the input is then connected to the residual. Perform the Hadamard product to obtain the final output value. .
2. The low-resolution industrial image classification method based on label-guided diversity contrastive learning as described in claim 1, characterized in that: The feature extractor uses a convolutional neural network structure, and the classifier uses a feedforward neural network structure, which is used to decode the representation vector from the feature extractor layer by layer, thereby achieving the classification of low-resolution images.
3. The low-resolution industrial image classification method based on label-guided diversity contrastive learning as described in claim 1, characterized in that: The feature extractor uses ResNet18 as the backbone, and the classifier consists of three layers: input layer, hidden layer, and output layer.
4. The low-resolution industrial image classification method based on label-guided diversity contrastive learning as described in claim 3, characterized in that: Final output value The formula is: ； The classifier includes a dropout layer to apply regularization constraints.
5. The low-resolution industrial image classification method based on label-guided diversity contrastive learning as described in any one of claims 1-4, characterized in that: The input to the feature extractor is the original low-resolution image. and their corresponding category labels The output is a representation vector containing discriminative information. The parameters of the neural network are optimized by minimizing the distance between the representation vectors of similar images and maximizing the distance between the representation vectors of dissimilar images. .
6. The low-resolution industrial image classification method based on label-guided diversity contrastive learning as described in any one of claims 1-4, characterized in that: The label-guided diversity contrastive learning strategy consists of two parts: label-guided contrastive learning and diversity contrastive learning. Label-guided contrastive learning learns a low-dimensional representation by maximizing the consistency of positive sample pairs and minimizing the consistency of negative sample pairs. ; Let the dataset There are a total of n samples Then the loss function for label-guided contrastive learning is: in, and Let them represent the positive sample pair set and the negative sample pair set, respectively. These are positive sample pairs of sample x. These are negative sample pairs of sample x. It is the mapping relationship learned by the neural network.
7. The low-resolution industrial image classification method based on label-guided diversity contrastive learning as described in claim 6, characterized in that: Diversity contrastive learning introduces an additional quadratic regularization term based on the Gram matrix into the loss function of label-guided contrastive learning. To penalize similarity and promote orthogonalization of the weight matrix W to learn more diverse representations, the loss function for label-guided diversity contrastive learning is: 。 8. A low-resolution industrial image classification system based on label-guided diversity contrastive learning, characterized in that: The system includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it performs the low-resolution industrial image classification method based on label-guided diversity contrastive learning as described in any one of claims 1-7.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions that, when invoked and executed by a processor, cause the processor to implement the low-resolution industrial image classification method based on label-guided diversity contrastive learning as described in any one of claims 1-7.