Regional feature fuzzy-guided image ordered regression prediction method and system

Abstract

The invention discloses an image ordered regression prediction method and system based on fuzzy guidance of regional features. The method comprises the following steps: generating a fine-grained region-level pseudo tag for a to-be-graded image by using an offline region-level labeling module; after to-be-graded image features are extracted, region-level and channel-level ordered label fuzzy learning is performed on the extracted image features based on fine-grained region-level pseudo labels through a two-level fuzzy feature learning module, and learned fuzzy image representation is obtained. And representing the prediction grading probability distribution based on the blurred image as an ordered label prediction result. The method and the system are used for improving the accuracy of feature-based hierarchical boundaries and processing fuzzy ordered labels, and are particularly suitable for image classification tasks requiring detail sensitive decisions.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular, relates to an image ordered regression prediction method and system guided by regional feature fuzziness. Background Art

[0002] Image ordered regression lies at the intersection of two fundamental paradigms, classification and regression, and aims to predict the corresponding ordered label for an input image. This task not only requires the model to understand the content of the image but also to capture the inherent order relationship between class labels. For example, in application scenarios such as facial age estimation, image aesthetics evaluation, historical image dating, and medical disease grading, image ordered regression methods have shown significant value.

[0003] The goal of image ordered regression is to learn a mapping rule to assign the input image to a specific ordered rank. Traditional ordered regression methods mainly follow the regression or classification paradigm and adopt traditional optimization objectives such as mean absolute / squared error or cross-entropy loss. However, these methods do not fully utilize the unique property of ordered regression, that is, the labels have an order. To address this property, many methods have transformed the direct ordered label prediction problem through different strategies and effectively utilized the ordered information.

[0004] In addition, another challenge faced by image ordered regression is the ambiguity of label definition. Weak class distinguishability, human annotator heterogeneity, and external factors can all lead to the ambiguity of observed labels, which will affect the model's ability to fit the data. Some existing studies have addressed this challenge by using the label distribution learning paradigm, which trains the model through the label distribution of instances. In addition, DLDL transforms each image label into a discrete label distribution and learns the label distribution by minimizing the KL divergence between the predicted label distribution and the true label distribution. The OLDL model further incorporates the ordered property of labels into the LDL paradigm, based on spatial, semantic, and temporal order relationships. Another research direction focuses on adjusting the feature space according to label ambiguity. In this context, some methods use probabilistic embeddings to represent each sample as a Gaussian distribution instead of a fixed point for classification. In addition, POEs proposes an ordered distribution constraint to maintain the ordered relationship in the latent space.

[0005] In recent years, the combination of fuzzy systems and deep learning has shown significant potential in image classification tasks, especially in handling uncertainty and improving the accuracy of various applications. This hybrid approach not only enhances image classification accuracy through advanced feature extraction models but also addresses challenges such as boundary point classification, providing better knowledge representation and reasoning capabilities under uncertainty. In image classification, the integration of fuzzy logic and CNN has proven effective in specific tasks, such as in colon polyp detection, where it has achieved remarkable performance by improving feature learning. Additionally, fuzzy deep models based on fuzzy restricted Boltzmann machines (RBMs) have been proposed for high-dimensional data classification, especially in scenarios involving raw high-dimensional images. These methods outperform traditional methods in terms of classification accuracy. Furthermore, advanced architectures that combine CNN with fuzzy clustering techniques, such as fuzzy clustering and fuzzy C-means clustering, have been shown to be superior to traditional methods in classification accuracy. Summary of the Invention

[0006] In view of the above, the object of the present invention is to provide a method and system for image ordinal regression prediction guided by regional feature fuzziness. As a new technical solution, it involves the image ordinal regression technology in computer vision, aiming to improve the accuracy of feature-based hierarchical boundaries and handle fuzzy ordinal labels, and is particularly applicable to image classification tasks that require detail-sensitive decisions, such as the automatic grading of diabetic retinopathy.

[0007] To achieve the above object of the invention, an embodiment provides a method for image ordinal regression prediction guided by regional feature fuzziness, which includes the following steps:

[0008] Use an offline regional-level annotation module to generate fine-grained regional-level pseudo-labels for the image to be graded;

[0009] After extracting the features of the image to be graded, through a two-level fuzzy feature learning module, perform regional-level and channel-level ordinal label fuzziness learning on the extracted image features based on the fine-grained regional-level pseudo-labels, and obtain the learned blurred image representation. Predict the grading probability distribution based on the blurred image representation as the ordinal label prediction result.

[0010] Preferably, using the offline regional-level annotation module to generate fine-grained regional-level pseudo-labels for the image to be graded includes:

[0011] The offline regional-level annotation module includes a regional-level pseudo-label generation model enhanced by adjacent category mixing augmentation training. Use the regional-level pseudo-label generation model to offline generate fine-grained regional-level pseudo-labels for the image to be graded;

[0012] The regional-level pseudo-label generation model includes a backbone network encoder for extracting image features and a first regional-level classifier for generating regional-level pseudo-labels based on the extracted features.

[0013] Preferably, enhancing the sample to enhance the training of the regional-level pseudo-label generation model through adjacent category mixing, including:

[0014] Collect sample images and corresponding labels at each level. For each pair of image samples of adjacent categories, use the backbone network encoder to extract image features respectively, and perform vector addition mixing on the image features according to certain weights to obtain a new sample representation. The label corresponding to this new sample representation is calculated by mixing the original global true labels of the image sample pair according to the same weighted addition. Use the new sample representation and its label to enhance the training of the regional-level classifier to improve the discrimination ability of the regional-level classifier between boundary categories.

[0015] Preferably, perform ordered label ambiguity learning at the regional level and channel level respectively on the extracted image features based on the fine-grained regional-level pseudo-labels through a two-level fuzzy feature learning module and obtain the learned image blurred representation, including:

[0016] Separate the image regional-level features and channel-level features from the extracted image features, and capture the uncertainty in the image blur feature-label relationship from these two perspectives of image regional-level features and channel-level features, that is, calculate the membership grade and activation intensity through the Gaussian membership function and the AND fuzzy logic operation, and capture the mutual relationship between each sub-region or channel, so as to generate the image blurred representation f 1 and f 2 ;

[0017] Based on the two image blurred representations f 1 and f 2 and introduce the fine-grained regional-level pseudo-labels and true grading labels to perform co-training of the graded reliable regions and graded unreliable regions to achieve ordered label ambiguity learning and update the fuzzy learning parameters. Synthesize the learned image blurred representations f 1 and f 2 to obtain the learned image blurred representation.

[0018] Preferably, the process of generating the image blurred representation f 1 from the perspective of the image regional-level features by capturing the uncertainty in the image blur feature-label relationship is:

[0019] Adopt a set of Gaussian membership functions to convert the input image regional-level feature H k into the corresponding membership degree m is the group index of the Gaussian membership function, corresponding to each of the fuzzy rules, k is the index of the sub-region, μ mk and Represent the mean and variance of the Gaussian function to be learned and updated, aggregate these membership grades through the AND fuzzy logic operation, and calculate the activation strength of each fuzzy rule. Thus, capture the mutual relationships between different sub-regions, and obtain the fuzzy image representations corresponding to l1 fuzzy rules. Among them, d represents the channel dimension, and K is the total number of sub-regions, that is, the region dimension.

[0020] Generate the image fuzzy representation f from the perspective of channel-level features. 2 The calculation process of 1 is the same as that of generating the image fuzzy representation f from the perspective of image region-level features. The difference is that the input is channel-level features, and the corresponding fuzzy image representation is generated. Among them, l2 represents the number of fuzzy rules from the channel perspective.

[0021] Preferably, based on the two image fuzzy representations f 1 and f 2 and introduce fine-grained region-level pseudo-labels and true grading labels to perform collaborative training of graded reliable regions and graded unreliable regions to realize ordered label ambiguity learning and update the fuzzy learning parameters, including:

[0022] Construct training frameworks A and B. Each training framework includes a two-level fuzzy feature learning module, a second region-level classifier, a global classifier, and a reliable region feature filter. Based on these two training frameworks, perform ordered label ambiguity learning to update the fuzzy learning parameters. The specific process is as follows:

[0023] For the current training round:

[0024] In training framework A, for the image fuzzy representation f of each image 1 Based on the global grading prediction result of the global classifier, calculate the cross-entropy loss between this global grading prediction result and the global true label. At the same time, for the channel fuzzy image representation f 2 Based on the region grading prediction result of the second region-level classifier, calculate the cross-entropy loss between this region grading prediction result and the locally generated pseudo-label. And in the reliable region feature filter, model the cross-entropy loss distribution between the region grading prediction result and the locally generated pseudo-label to determine the credibility probability of each region. By setting the threshold hyperparameter τ, construct the region-level mask matrix M, and perform mask processing on the credibility probability of each region based on the constructed region-level mask matrix M to distinguish graded reliable regions and graded unreliable regions.

[0025] In training framework B, for the graded reliable regions determined based on training framework A, simultaneously utilize f 1 and f 2The prediction results are used to calculate the cross-entropy with the global true labels and regional pseudo-labels to update the fuzzy learning parameters; for the unreliable region grading determined based on Training Framework A, the regenerative pseudo-labels are calculated based on the region grading prediction results obtained from the current round of Training Framework A and the region grading prediction results obtained from the previous round of Training Framework B, and the fuzzy learning parameters are updated based on the mean square error between the regenerative pseudo-labels and the region grading prediction results, where specifically, the average set of the region grading prediction results obtained from the current round of Training Framework A and the region grading prediction results obtained from the previous round of Training Framework B is used as the regenerative pseudo-labels;

[0026] For the next training round:

[0027] Training Frameworks A and B are swapped, that is, Training Framework B is used as Training Framework A, and Training Framework A is used as Training Framework B, and then the process of the current training round is executed.

[0028] Preferably, calculating the regenerative pseudo-labels based on the region grading prediction results obtained from the current round of Training Framework A and the region grading prediction results obtained from the previous round of Training Framework B includes:

[0029] Using the average set of the region grading prediction results obtained from the current round of Training Framework A and the region grading prediction results obtained from the previous round of Training Framework B as the regenerative pseudo-labels.

[0030] To achieve the above invention purpose, the embodiment also provides a region feature fuzzy-guided image ordered regression prediction system, including:

[0031] A pseudo-label generation unit, which is used to generate fine-grained region-level pseudo-labels for the image to be graded by using the offline region-level annotation module;

[0032] A fuzzification learning and grading prediction unit, which is used to extract the features of the image to be graded, and then through the dual-level fuzzy feature learning module, perform ordered label fuzziness learning at the region level and channel level on the extracted image features based on the fine-grained region-level pseudo-labels, and obtain the fuzzified image representation after learning, and predict the grading probability distribution based on the fuzzified image representation as the ordered label prediction result.

[0033] To achieve the above invention purpose, the embodiment also provides a computing device, including a memory and one or more processors, where executable code is stored in the memory, and when the one or more processors execute the executable code, it is used to implement the above region feature fuzzy-guided image ordered regression prediction method.

[0034] To achieve the above invention purpose, the embodiment also provides a computer-readable storage medium, on which a program is stored, and when the program is executed by a processor, it implements the above region feature fuzzy-guided image ordered regression prediction method.

[0035] Compared with the prior art, the beneficial effects of the present invention at least include:

[0036] (1) The present invention can generate fine-grained regional pseudo-labels through the offline regional-level annotation module with only global ordered labels. This mechanism can explicitly capture the regional features affecting the hierarchical labels, significantly improving the attention to local details, thereby enhancing the classification accuracy;

[0037] (2) The present invention quantitatively analyzes label ambiguity through the fuzzy representation learning mechanism, effectively addressing the inherent label ambiguity and sequential challenges in the image ordinal regression task. Through the membership degree calculation and aggregation method, the model can more accurately handle the uncertainty between category boundaries, obtain a blurred image representation, and perform image ordinal regression prediction based on this, improving the accuracy of hierarchical prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0039] Figure 1 is a flowchart of the image ordinal regression prediction method guided by regional feature fuzziness provided by the embodiment;

[0040] Figure 2 is a structural schematic diagram of the offline regional-level annotation module and the dual-level fuzzy feature learning module provided by the embodiment;

[0041] Figure 3 is a structural schematic diagram of the training framework provided by the embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] To make the objectives, technical solutions, and advantages of the present invention clearer, the present 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 only used to explain the present invention and do not limit the protection scope of the present invention.

[0043] As Figure 1 shown, the embodiment provides an image ordinal regression prediction method guided by regional feature fuzziness, including the following steps:

[0044] S1, using the offline regional-level annotation module to generate fine-grained regional pseudo-labels for the image to be classified.

[0045] As Figure 2As shown in (a), the offline region-level annotation module includes a region-level pseudo-label generation model enhanced by adjacent-class mixed augmentation samples. The region-level pseudo-label generation model includes a backbone network encoder for extracting image features and a first region-level classifier for generating region-level pseudo-labels based on the extracted features. Among them, a pre-trained visual model is used as the backbone network encoder, which encodes the input image into embedded image features H ∈ R K×d , where K is the number of partition regions and d is the embedding dimension. The first region-level classifier can use a fully connected layer and a convolutional layer (such as a convolutional kernel of 1×1) as the classification head to calculate the embedded image features H to generate fine-grained region-level pseudo-labels.

[0046] To improve the ability of the region-level pseudo-label generation model to capture the differences between adjacent-class samples, an adjacent-class mixing method is introduced. Through adjacent-class mixed augmentation samples for training, that is, by controllably mixing the sample features of adjacent classes, new samples and corresponding mixed labels are created, and these augmented data are used to train the region-level pseudo-label generation model. The specific process is as follows:

[0047] Collect sample images and corresponding labels at each level. For each pair of image samples of adjacent classes, use the backbone network encoder to extract image features respectively, and perform vector addition mixing on the image features according to certain weights to obtain a new sample representation. The label corresponding to the new sample representation is calculated by performing the same weighted addition mixing on the original true global labels of the image sample pair. Use the new sample representation and its label to enhance the training of the region-level classifier to improve the discrimination ability of the region-level classifier between boundary classes.

[0048] After completing the enhanced training, the offline region-level annotation module enters the inference stage. At this time, instead of relying on the true global true labels, based on the learned knowledge, use the trained first region-level classifier to generate fine-grained region-level pseudo-labels for each region of each input image, providing an additional supervision signal for subsequent ordered label ambiguity learning, so that the regional features affecting the grading labels can be explicitly captured during fuzzy feature learning, thereby improving the accuracy when dealing with fuzzy class boundaries.

[0049] S2. After extracting the image features to be graded, the two-level fuzzy feature learning module performs ordered label ambiguity learning at the region level and channel level respectively on the extracted image features based on the fine-grained region-level pseudo-labels and obtains the learned blurred image representation. Predict the grading probability distribution based on the blurred image representation as the ordered label prediction result.

[0050] The two-level fuzzy feature learning module considers the influence of class ambiguity on features in different spatial position regions or channel features, and uses a set of Gaussian membership functions to convert the input region features into corresponding membership grades, which reflects the degree of association of each region or channel with different fuzzy linguistic term labels. Subsequently, all membership degrees are integrated through fuzzy logic to calculate the activation intensity of fuzzy rules. This process decomposes the modeling process of label ambiguity. First, the features are fuzzified through membership degree calculation, and then the fuzzy feature-label relationship is learned based on the aggregated fuzzy rules. Finally, a transformation space is constructed through non-linear projection mapping, enabling the model to better adapt to the fuzzy definition of labels in the ordinal regression task.

[0051] In the embodiment, the backbone network encoder is also used to extract the image features to be graded, and then through as Figure 2 shown in (b) of the two-level fuzzy feature learning module, the two-level fuzzy feature learning module performs ordered label ambiguity learning at the region level and channel level on the extracted image features based on the fine-grained region-level pseudo-labels and obtains the learned image fuzzified representation, specifically including:

[0052] (a) The image region-level features and channel-level features are separated from the extracted image features, and the uncertainty in the image fuzzified feature-label relationship is captured from the two perspectives of the image region-level features and channel-level features to quantitatively analyze the label ambiguity, that is, the membership grade and activation intensity are calculated through Gaussian membership functions and AND fuzzy logic operations, and the mutual relationship between each sub-region or channel is captured, so as to generate the image fuzzified representation f 1 and f 2 .

[0053] During the embodiment process, the execution formulas of the two perspectives are similar. Taking the region-level fuzzification process as an example, a set of Gaussian membership functions is used to convert the input image region-level features H k into the corresponding membership degrees m is the group index of the Gaussian membership function, corresponding to each fuzzy rule, k is the index of the sub-region, μ mk and represent the mean and variance of the Gaussian function to be learned and updated. These membership grades are aggregated through AND fuzzy logic operations to calculate the activation intensity thus capturing the mutual relationship between different sub-regions and obtaining the fuzzified image representation corresponding to l1 fuzzy rules where d represents the channel dimension, K is the total number of sub-regions, that is, the region dimension, and the aggregation can adopt the Concat operation;

[0054] For channel-level fuzzification, the image fuzzified representation f from the perspective of channel-level features is generated2 The calculation process is the same as the image blurring representation f from the perspective of generating image region-level features 1 The difference is that the input is channel-level features, and the aggregation direction is in the region direction, corresponding to generating the blurring image representation Among them, l2 represents the number of blurring rules from the channel perspective.

[0055] (b) Based on two image blurring representations f 1 and f 2 And introduce fine-grained region-level pseudo-labels and true grading labels to perform co-training of grading reliable regions and grading unreliable regions to realize ordered label ambiguity learning and update the fuzzy learning parameters. Specifically, an online co-training strategy is adopted to perform additional attention on valuable image regions. In order to reduce the influence of pseudo-label noise and provide additional supervision information, this strategy uses two training frameworks to train alternately and provide each other with the selection mask matrix of high-confidence regions. This strategy uses a Gaussian mixture model to model the cross-entropy loss distribution and determine the credibility probability of each region. Then, based on a preset threshold, an image region mask matrix is constructed to distinguish reliable and unreliable regions. For reliable regions, the pseudo-labels and the new prediction probabilities of the model are updated by linear combination; for unreliable regions, the average set of the predictions of the two models is used as the regenerated pseudo-labels to ensure that the model can effectively learn valuable regional features. The specific process includes:

[0056] Construct training framework A and training framework B. Each training framework has the same structure, as Figure 3 shown, both contain a two-level fuzzy feature learning module, a second region-level classifier, a global classifier, and a reliable region feature filter. Based on the co-training strategy of these two training frameworks, ordered label ambiguity learning is performed to update the fuzzy learning parameters.

[0057] For the current training round:

[0058] In training framework A, for the image blurring representation f of each image 1 Based on the global grading prediction result of the global classifier, calculate the cross-entropy loss between the global grading prediction result and the global true label. At the same time, for the channel blurring image representation f 2 Based on the region grading prediction result of the second region-level classifier, calculate the cross-entropy loss between the region grading prediction result and the locally generated pseudo-label, and in the reliable region feature filter, model the cross-entropy loss distribution between the region grading prediction result and the locally generated pseudo-label to determine the credibility probability of each region. By setting the threshold hyperparameter τ, construct the region-level mask matrix M, and perform mask processing on the credibility probability of each region based on the constructed region-level mask matrix M to distinguish grading reliable regions and grading unreliable regions;

[0059] In training framework B, for the hierarchical reliable regions determined based on training framework A, the cross-entropy is calculated using the prediction results of f 1 and f 2 with the global ground truth label and the regional pseudo-label to update the fuzzy learning parameters; for the hierarchical unreliable regions determined based on training framework A, the regenerated pseudo-label is calculated based on the regional hierarchical prediction result obtained from the current round of training framework A and the regional hierarchical prediction result obtained from the previous round of training framework B, and the fuzzy learning parameters are updated based on the mean square error between the regenerated pseudo-label and the regional hierarchical prediction result, where specifically, the average set of the regional hierarchical prediction result obtained from the current round of training framework A and the regional hierarchical prediction result obtained from the previous round of training framework B is used as the regenerated pseudo-label;

[0060] For the next training round:

[0061] Training frameworks A and B are exchanged, that is, training framework B is used as training framework A, and training framework A is used as training framework B, and then the process of the current training round is executed above.

[0062] Through this mutual supervision mechanism, the two training frameworks are continuously optimized to ensure that the model can not only accurately predict the global label but also effectively utilize the regional-level pseudo-label for fine-grained learning, reduce the influence of noisy pseudo-labels and pay extra attention to the regions with high influence on the hierarchical prediction results, enhance the robustness and generalization ability of the mean and variance of the Gaussian function in fuzzy feature learning, and ensure the effective learning of fine-grained features and high-quality ordered classification performance.

[0063] After the above learning, the comprehensively learned image fuzzified representation f 1 and f 2 are used to obtain the learned image fuzzified representation, and then the hierarchical probability distribution is predicted based on the fuzzified image representation as the ordered label prediction result. Specifically, a convolutional layer (such as a convolutional kernel of 1×1) and a linear layer are used as the classification head to provide the probability of predicting the ordered label.

[0064] The method provided by the embodiments of the present invention can effectively address the limitations of the prior art in dealing with label uncertainty and orderliness, significantly improve the performance of the image ordered regression task, improve the classification accuracy and the learning ability of local features, and is particularly suitable for application scenarios that require detail-sensitive decisions.

[0065] When the above region feature fuzzy-guided image ordered regression prediction method is used for the automatic grading of diabetic retinopathy, an image ordered regression dataset DR is prepared for the offline training of the region-level pseudo-label generation model and the ordered label fuzziness learning of the two-level fuzzy feature learning module.

[0066] DR is a dataset specifically for the task of automatic grading of diabetic retinopathy, aiming to help researchers and medical professionals develop and validate algorithms that can accurately assess the severity of retinopathy in fundus images. Diabetic retinopathy is one of the common complications in diabetic patients. Early detection and accurate grading are crucial for timely treatment and prevention of vision loss. This dataset contains 35,126 retinal fundus photographs, and each image is labeled by professional doctors into one of five levels, representing different degrees of retinopathy: No DR (no lesion), Mild DR (mild lesion), Moderate DR (moderate lesion), Severe DR (severe lesion), and Proliferative DR (proliferative lesion).

[0067] A notable feature of this dataset is its highly imbalanced class distribution. Approximately 73.5% of the samples are labeled as "no lesion", while the number of samples in the other four levels is relatively small. This imbalance poses challenges to model training and appropriate strategies need to be adopted to balance the influence of different classes, such as oversampling, undersampling, or using weighted loss functions. In addition, different levels of retinopathy may only show subtle differences in the images, so the model needs to have the ability to capture key lesion areas, such as bleeding points, hard exudates, microaneurysms, etc., which places higher requirements on the model's fine-grained feature extraction ability.

[0068] To ensure the consistency and standardization of the input, all images were first preprocessed, including resizing, normalization, and necessary augmentation to improve the generalization ability and robustness of the model. According to the labeled data, the dataset was divided according to a predetermined ratio (e.g., 70% for training and 30% for validation). Importantly, when dividing, it is necessary to ensure that the proportion of samples in each category remains consistent to avoid model bias caused by class imbalance.

[0069] The embodiment also provides an image ordered regression prediction system guided by regional feature fuzziness, including: a pseudo-label generation unit and a fuzzification learning and grading prediction unit. The pseudo-label generation unit is used to generate fine-grained regional-level pseudo-labels for the image to be graded by using an offline regional-level annotation module; the fuzzification learning and grading prediction unit is used to extract the features of the image to be graded, and then through a two-level fuzzy feature learning module, perform regional-level and channel-level ordered label fuzziness learning on the extracted image features based on the fine-grained regional-level pseudo-labels to obtain the learned fuzzified image representation, and predict the grading probability distribution based on the fuzzified image representation as the ordered label prediction result.

[0070] It should be noted that when the above-mentioned method and device for image ordinal regression prediction guided by regional feature blurring perform image ordinal regression prediction, the above-mentioned functional modules or units can be used for illustration. The above functions can be allocated to different functional modules or units according to needs, that is, the internal structure of the terminal or server is divided into different functional modules or units to complete all or part of the functions described above. In addition, the above-mentioned system for image ordinal regression prediction guided by regional feature blurring and the embodiments of the method for image ordinal regression prediction guided by regional feature blurring belong to the same concept. For the specific implementation process, please refer to the embodiments of the method for image ordinal regression prediction guided by regional feature blurring, which will not be elaborated here.

[0071] Based on the same inventive concept, the embodiment also provides a computing device, including a memory and one or more processors. The memory stores executable code. When the one or more processors execute the executable code, they are used to implement the above-mentioned method for image ordinal regression prediction guided by regional feature blurring, specifically including the following steps:

[0072] S1, using an offline region-level annotation module to generate fine-grained region-level pseudo-labels for the image to be classified;

[0073] S2, after extracting the features of the image to be classified, through a two-level fuzzy feature learning module, perform region-level and channel-level ordinal label fuzziness learning on the extracted image features based on the fine-grained region-level pseudo-labels, and obtain the blurred image representation after learning. Based on the blurred image representation, predict the classification probability distribution as the ordinal label prediction result.

[0074] , In terms of hardware, the computing device provided in the embodiment, in addition to including a processor and a memory, also includes other hardware required for other services such as an internal bus, a network interface, and a memory. The memory is a non-volatile memory. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the method for image ordinal regression prediction guided by regional feature blurring described in S1-S2 above. Of course, in addition to the software implementation method, the present invention does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, and can also be hardware or a logic device.

[0075] Based on the same inventive concept, the embodiment also provides a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the above-mentioned method for image ordinal regression prediction guided by regional feature blurring, specifically including the following steps:

[0076] S1, using an offline region-level annotation module to generate fine-grained region-level pseudo-labels for the image to be classified;

[0077] After extracting the features of the image to be classified, the two-level fuzzy feature learning module performs ordered label ambiguity learning at the regional level and the channel level on the extracted image features based on the fine-grained region-level pseudo-labels, and obtains the blurred image representation after learning. Based on the blurred image representation, the classification probability distribution is predicted as the ordered label prediction result.

[0078] In an embodiment, the computer-readable medium includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data.

[0079] The specific embodiments described above have elaborated on the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not used to limit the present invention. Any modifications, supplements, equivalent replacements, etc. made within the scope of the principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. An image ordered regression prediction method guided by regional feature fuzziness, characterized in that Including the following steps: Using an offline region-level annotation module to generate fine-grained region-level pseudo-labels for the image to be classified; After extracting the features of the image to be classified, through a two-level fuzzy feature learning module, perform region-level and channel-level ordered label ambiguity learning on the extracted image features based on the fine-grained region-level pseudo-labels, and obtain the blurred image representation after learning. Predict the classification probability distribution based on the blurred image representation as the ordered label prediction result.

2. The method for predicting the ordered regression of an image with fuzzy guidance of regional features according to claim 1, wherein, Using an offline region-level annotation module to generate fine-grained region-level pseudo-labels for the image to be classified, including: The offline region-level annotation module includes a region-level pseudo-label generation model enhanced by adjacent class mixed augmentation samples. Use the region-level pseudo-label generation model to offline generate fine-grained region-level pseudo-labels for the image to be classified; The region-level pseudo-label generation model includes a backbone network encoder for extracting image features and a first region-level classifier for generating region-level pseudo-labels based on the extracted features.

3. The method for predicting image ordered regression with fuzzy guidance of regional features according to claim 2, wherein Enhancing the training of the region-level pseudo-label generation model through adjacent class mixed augmentation samples, including: Collect sample images at each level and their corresponding labels. For each pair of adjacent class image samples, use the backbone network encoder to extract image features respectively, and perform vector addition mixing on the image features with certain weights to obtain a new sample representation. The label corresponding to the new sample representation is calculated by performing the same weighted addition mixing on the original global true labels of the image sample pair. Use the new sample representation and its label to enhance the training of the region-level classifier to improve the discrimination ability of the region-level classifier between boundary classes.

4. The method for predicting the ordered regression of an image with fuzzy guidance of regional features according to claim 1, wherein Performing region-level and channel-level ordered label ambiguity learning on the extracted image features based on the fine-grained region-level pseudo-labels through a two-level fuzzy feature learning module and obtaining the blurred image representation after learning, including: Separate the image region-level features and channel-level features from the extracted image features, and capture the uncertainty in the image blurring feature-label relationship from these two perspectives of the image region-level features and channel-level features. That is, calculate the membership grade and activation intensity through the Gaussian membership function and the AND fuzzy logic operation, capture the mutual relationship between each sub-region or channel, so as to generate the image blurring representation f from the perspectives of the image region-level features and channel-level features 1 and f 2 ; Based on two image blurring representations f 1 and f 2 And introduce fine-grained region-level pseudo-labels and true hierarchical labels to perform collaborative training on hierarchical reliable regions and hierarchical unreliable regions to achieve ordered label ambiguity learning, update the fuzzy learning parameters, and comprehensively learn the image blurring representations f 1 and f 2 Obtain the learned image blurring representation.

5. The method for predicting the ordered regression of an image with fuzzy guidance of regional features according to claim 4, wherein From the perspective of the image region-level features, the process of generating the image blurring representation f by capturing the uncertainty in the image blurring feature-label relationship 1 is as follows: Convert the input image region level feature H using a set of Gaussian membership functions k to the corresponding membership degrees where m is the group index of the Gaussian membership functions, corresponding to each of the fuzzy rules, k is the index of the sub-region, μ mk and represent the mean and variance of the Gaussian function to be learned and updated. Aggregate these membership grades through AND fuzzy logic operations to calculate the activation strength of each fuzzy rule thus capturing the interrelationships between different sub-regions and obtaining the fuzzy image representations corresponding to l1 fuzzy rules where d represents the channel dimension, K is the total number of sub-regions, i.e., the region dimension; Generate the image blurring representation f from the perspective of channel-level features 2 The calculation process of 1 is the same as that of generating the image blurring representation f from the perspective of image region-level features. The difference is that the input is channel-level features, corresponding to generating the blurring image representation where l2 represents the number of blurring rules from the channel perspective.

6. The method for predicting the ordered regression of an image with fuzzy guidance of regional features according to claim 4, wherein Based on two blurred image representations f 1 and f 2 And introducing fine-grained region-level pseudo-labels and true hierarchical labels for collaborative training of reliable and unreliable hierarchical regions to achieve ordered label ambiguity learning and update fuzzy learning parameters, including: Construct training frameworks A and B. Each training framework includes a two-level fuzzy feature learning module, a second region-level classifier, a global classifier, and a reliable region feature filter. Perform ordered label ambiguity learning based on these two training frameworks to update the fuzzy learning parameters. The specific process is as follows: For the current training round: In training framework A, for the image blurring representation f of each image 1 Based on the global grading prediction result of the global classifier, calculate the cross-entropy loss between the global grading prediction result and the global true label. At the same time, for the channel blurring image representation f 2 Based on the regional grading prediction result of the second regional-level classifier, calculate the cross-entropy loss between the regional grading prediction result and the locally generated pseudo-label. In the reliable region feature filter, model the cross-entropy loss distribution between the regional grading prediction result and the locally generated pseudo-label to determine the credibility probability of each region. By setting the threshold hyperparameter τ, construct the regional-level mask matrix M, and perform mask processing on the credibility probability of each region based on the constructed regional-level mask matrix M to distinguish the grading reliable regions and the grading unreliable regions; In training framework B, for the hierarchical reliable regions determined based on training framework A, the cross-entropy is calculated using the prediction results of f 1 and f 2 with the global true label and the regional pseudo-label to update the fuzzy learning parameters; for the hierarchical unreliable regions determined based on training framework A, the regeneration pseudo-label is calculated based on the regional hierarchical prediction results obtained from the current round of training framework A and the regional hierarchical prediction results obtained from the previous round of training framework B, and the fuzzy learning parameters are updated based on the mean square error between the regeneration pseudo-label and the regional hierarchical prediction results, where specifically, the average set of the regional hierarchical prediction results obtained from the current round of training framework A and the regional hierarchical prediction results obtained from the previous round of training framework B is used as the regeneration pseudo-label; For the next training round: Exchange training frameworks A and B, that is, use training framework B as training framework A and training framework A as training framework B, and then execute the process of the current training round above.

7. The method for predicting the ordered regression of an image with fuzzy guidance of regional features according to claim 6, wherein Calculating the regenerated pseudo-labels based on the region classification prediction results obtained from training framework A in the current round and the region classification prediction results obtained from training framework B in the previous round, including: Taking the average set of the region classification prediction results obtained from training framework A in the current round and the region classification prediction results obtained from training framework B in the previous round as the regenerated pseudo-labels.

8. An image ordered regression prediction system guided by regional feature fuzziness, characterized in that, Including: A pseudo-label generation unit, which is used to generate fine-grained region-level pseudo-labels for the image to be classified by using an offline region-level annotation module; The fuzzification learning and hierarchical prediction unit is used to extract the features of the image to be graded, and then, through the dual-level fuzzy feature learning module, perform ordered label fuzziness learning at the regional level and channel level respectively on the extracted image features based on the fine-grained region-level pseudo-labels, and obtain the fuzzified image representation after learning. Based on the fuzzified image representation, the grading probability distribution is predicted as the ordered label prediction result.

9. A computing device, comprising a memory and one or more processors, wherein executable code is stored in the memory, characterized in that, When the one or more processors execute the executable code, they are used to implement the region feature fuzzy-guided image ordered regression prediction method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A program is stored thereon, and when the program is executed by the processor, it implements the region feature fuzzy-guided image ordered regression prediction method according to any one of claims 1-7.

Images