A fine-grained data augmentation method based on saliency map guidance

Abstract

This invention discloses a fine-grained data augmentation method based on saliency map guidance, comprising the following steps: First, generating a Grad-CAM saliency map of the source image; second, generating a mask for the source image based on the Grad-CAM saliency map, and then extracting foreground information from the source image based on the mask; generating a saliency map using the Grad-CAM method, using the foreground information extracted from the source image as a patch, and synthesizing the target image in the Grad-CAM region to generate the final image. Simultaneously, the semantic percentage maps of the source image region and the target image region in the augmented image are summed to obtain a new semantic label ratio between the source and target images in the augmented image, rather than directly using the area ratio of the source and target images. The fine-grained data augmentation method based on saliency map guidance proposed in this invention can select foreground in the source image for generating the augmented image through the guidance of Grad-CAM saliency map, avoiding the introduction of invalid regions and improving performance to a certain extent.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and more particularly to a fine-grained data augmentation method based on saliency map guidance. Technical Background

[0002] In the development of artificial intelligence, computer vision has become a crucial branch, especially image recognition and classification technology. Computer vision image recognition tasks refer to using computer algorithms to identify and classify objects, scenes, faces, etc., in images. However, image recognition tasks also face several challenges, one of which is data scarcity. Data scarcity refers to a situation where the amount of labeled data available for a given task and category is relatively small or insufficient. When the amount of available data is limited, models are prone to overfitting the training data, over-memorizing noise or specific features of samples. This leads to poor model performance on new samples and a lack of robustness. Data scarcity also makes annotation difficult. Labeling large-scale datasets typically requires significant time, manpower, and expertise. For certain specific domains or fine-grained tasks, due to the complexity of labeling specific categories, it is difficult to obtain large-scale, accurate labeled data.

[0003] In existing technologies, driven by deep learning, fine-grained image classification has made some progress. However, due to its complexity, its performance still falls far short of the needs of practical applications. This complexity stems primarily from two aspects: firstly, the differences between fine-grained categories are usually relatively small, possibly limited to subtle variations in shape, texture, or color; secondly, samples within the same category may exhibit numerous variations in morphology, pose, lighting, and scale, further increasing the difficulty of fine-grained image classification. Based on these premises, a fine-grained data augmentation method guided by saliency maps is proposed. Data augmentation can effectively address these problems. Data augmentation increases the number and diversity of samples by transforming or expanding existing data. For example, operations such as translation, rotation, scaling, and cropping can be performed to generate new training samples, increasing the robustness of the model. However, while traditional data augmentation methods can expand datasets, they often struggle to generate images that differ significantly from the original images while still retaining their category attributes, thus limiting their effectiveness in fine-grained classification tasks. Hybrid image data augmentation methods have become an important research direction in data augmentation techniques. By mixing two or more images in a certain way to generate new training samples, this method can effectively expand the diversity of the dataset and improve the generalization ability of the model.

[0004] Therefore, it is necessary to provide a saliency map-guided fine-grained data augmentation method to address the aforementioned issues. In fine-grained image classification problems, the differences between samples of the same class are usually small, while the differences between samples of different classes often manifest in subtle features, making fine-grained classification problems quite difficult. Hybrid image data augmentation methods can, to some extent, compensate for these differences and improve the model's performance on fine-grained classification problems. Summary of the Invention

[0005] The purpose of this invention is to provide a fine-grained data augmentation method based on saliency maps, thereby overcoming the shortcomings of existing technologies in terms of the difficulty of obtaining large-scale, accurately labeled data and the diversity of datasets in fine-grained image classification, and improving the generalization ability of the model.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a fine-grained data augmentation method based on saliency map guidance, effectively enhancing the sample confidence of model calibration. First, a source image is input, and a saliency map is generated using the Grad-CAM method. Second, a mask is generated for the source image based on the Grad-CAM saliency map. Third, foreground information of the source image is extracted based on the mask. For the target image, a saliency map is generated using the Grad-CAM method, and the foreground information extracted from the source image is used as a patch to synthesize the target image within the Grad-CAM region, completing the generation of the augmented image. Finally, the semantic proportions of the source and target images in the newly generated augmented image are calculated using the generated Grad-CAM saliency map, ultimately serving as the label for the augmented image. Since saliency maps help us understand the regions that the human visual system focuses on when observing images and provide information on the location and segmentation of salient regions in an image, this method, guided by the Grad-CAM saliency map, selects foreground regions in the source image for generating the augmented image, avoiding the introduction of background interference. The label of the newly generated augmented image is evaluated using Grad-CAM, making the augmented image more closely matched to its corresponding label.

[0007] This invention is achieved through the following technical solution: a fine-grained data augmentation method based on saliency maps, comprising the following steps:

[0008] Step 1: Input the source image and generate a saliency map using the lightweight method Grad-CAM;

[0009] Step 2: Generate a mask for the source image based on the Grad-CAM saliency map of the source image;

[0010] Step 3: Extract the foreground information of the source image based on the mask of the source image;

[0011] Step 4: Input the target image and generate a saliency map using the lightweight method Grad-CAM;

[0012] Step 5: Use the foreground information extracted from the source image as a patch to synthesize the Grad-CAM region of the target image;

[0013] Step 6: Calculate the new label for the final enhanced image.

[0014] As a preferred embodiment of the fine-grained data augmentation method based on saliency map guidance described in this invention, foreground regions with salient speech meaning in the source and target images are located. Based on this, data augmentation is achieved by replacing salient semantic information in the source and target images. The designed data augmentation method can accurately locate effective foreground information to achieve diversified data expansion, thereby improving the robustness of the data augmentation method.

[0015] As a preferred embodiment of the fine-grained data augmentation method based on saliency map guidance described in this invention, in step one, in order to obtain the Grad-CAM saliency map of category c... First, calculate the weight of the k-th feature map for category c.

[0016]

[0017] Where Z is the number of pixels in the feature map. y represents the weight of the k-th feature map. c Let A represent the score gradient of category c. k This is the feature map of the k-th channel after the last feature mapping layer of the convolutional layer. A represents k The (i,j)th feature point in the middle.

[0018] The final Grad-CAM saliency map is shown below:

[0019]

[0020] Where φ(·) represents the operation of upsampling along the feature map dimension to match the input image size. Let S represent the k-th saliency map of the source image, where ReLU is the activation function and S represents the source image.

[0021] As a preferred embodiment of the fine-grained data augmentation method based on saliency map guidance described in this invention, in step two, based on the saliency map of the source image... Generate a mask. Threshold the saliency map, selecting a threshold based on the pixel intensity to generate a binary mask.

[0022] As a preferred embodiment of the fine-grained data augmentation method based on saliency map guidance described in this invention, in step three, the foreground information of the source image is extracted based on the mask of the source image. For source image I... s Saliency map detection can be represented as:

[0023]

[0024] in, Representative source image I s The saliency plot, f Grad-CAM (·) represents the Grad-CAM saliency map generation model; it searches for the pixel with the maximum saliency value in the saliency map; M represents the mask matrix, which can be calculated using the following formula:

[0025]

[0026] Here, argmax(·) represents the parameter function that takes the maximum value.

[0027] As a preferred embodiment of the fine-grained data augmentation method based on saliency map guidance described in this invention, in step four, the target image is input, and a saliency map is generated using the lightweight method Grad-CAM; the final Grad-CAM saliency map is represented as follows:

[0028]

[0029] Where φ(·) represents the operation of upsampling along the feature map dimension to match the input image size. This represents the weight of the k-th feature map. Let A represent the k-th saliency map of the target image. k This represents the feature map of the last convolutional layer, where T represents the target image.

[0030] As a preferred embodiment of the fine-grained data augmentation method based on saliency map guidance described in this invention, in step five, the foreground information extracted from the source image is used as a patch for synthesis in the Grad-CAM region of the target image. The blending of the two images can be defined as follows:

[0031] I mix =M⊙I s +(1-M)⊙I t (6)

[0032] Among them, I mix Denotes a blended image, M∈{0,1} W×H ⊙ represents a binary mask, and ⊙ represents element-wise multiplication.

[0033] As a preferred embodiment of the saliency map-guided fine-grained data augmentation method described in this invention, in step six, the generated Grad-CAM saliency map is normalized, and the percentage map SPM of the image is calculated in the following manner:

[0034]

[0035] in This represents the Grad-CAM saliency map for category c.

[0036] The proportions of source and target images in the enhanced image labels are as follows:

[0037]

[0038]

[0039] Where M is the source image patch location mask in the enhanced image generation process. To increase the proportion of source images in image tags, To increase the proportion of the target image in the image labels, the final labels for the enhanced images are:

[0040]

[0041] Among them, y mix To enhance image labels.

[0042] The beneficial effects of this invention are as follows: This invention proposes a fine-grained data augmentation method based on saliency map guidance. By leveraging the location and semantic information of salient regions, it can expand the diversity of augmented images, helping the model better learn the foreground semantic information relevant to fine-grained categories. This improves the performance and robustness of fine-grained image recognition tasks by expanding the dataset. In fine-grained recognition, although the spatial distribution of image saliency information is extremely uneven, the proposed method can accurately locate salient semantic information, preventing the introduction of significant label noise into the augmented data. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] in:

[0045] Figure 1 This invention provides an overall framework diagram for a fine-grained data augmentation method based on saliency maps.

[0046] Figure 2 A flowchart illustrating a fine-grained data augmentation method based on saliency map guidance provided by this invention;

[0047] Figure 3 This invention provides a salient graph-guided fine-grained data augmentation method, illustrating the network performance of different training strategies compared to traditional image augmentation. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0050] Secondly, the present invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include three-dimensional spatial dimensions of length, width, and depth.

[0051] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0052] Example 1, as Figure 1 As shown, this invention proposes a fine-grained data augmentation method based on saliency map guidance. First, a Grad-CAM saliency map is generated. Given an image and the class of interest c as input, the image is forward-propagated through the CNN part of the model. Second, the raw score y of that class is obtained through task-specific computation. c Next, the gradient is backpropagated to the last convolutional layer, and the gradients for all classes except the desired class c (which is set to 1) are set to 0. Third, the gradient is multiplied by the corresponding feature map and summed over all feature maps to obtain the Grad-CAM saliency map. The generated Grad-CAM saliency map can be used not only to select foreground patches from the source image to generate enhanced images, but also as the label ratio between the source and target images in the newly generated hybrid image during hybrid label generation.

[0053] Figure 2 This is a flowchart of a fine-grained data augmentation method based on saliency map guidance provided by the present invention, including the following steps:

[0054] Step 1: Source image, generate saliency map using the lightweight Grad-CAM method;

[0055] To obtain the Grad-CAM saliency map of source image category c This paper first calculates the score gradient y for category c. c and the feature mapping A of the last convolutional layer k Correspondence The corresponding formula is as follows:

[0056]

[0057] Where X represents the input, f conv2d This represents the calculation of the convolutional layer. This represents the computation of the fully connected layer, where W1 represents the weights of the convolutional layer. This represents the weight corresponding to the first node in the fully connected layer. Here, let (f) conv2d The output of (X,W1) is O=(O 11 O 12 O 21 O 22 ), then y c The calculation formula is:

[0058]

[0059] Where O represents the vector calculated by the convolutional layer, and W represents the corresponding weight in the fully connected layer.

[0060] This is used to calculate the gradients of these backpropagations, which are then globally averaged pooled along the width and height dimensions (indexed by i and j, respectively) to obtain the weights of the k-th feature map for class c.

[0061]

[0062] Where Z is the number of pixels in the feature map. y represents the weight of the k-th feature map. c Let A represent the score gradient of category c. k This is the feature map of the k-th channel after the last feature mapping layer of the convolutional layer. A represents k The (i,j)th feature point. It can be seen from the calculation formula (10) that... It involves predicting the predicted score y for category c. cBackpropagation is performed, and then the gradient information backpropagated to feature layer A is used to calculate the importance of each channel k in feature layer A. Next, the data for each channel of feature layer A is weighted and summed using α. After obtaining the weights of the class on all feature maps, the weighted sums are calculated, a ReLU activation function is applied, and upsampling is performed to obtain the final Grad-CAM saliency map.

[0063]

[0064] Where φ(·) represents the operation of upsampling along the feature map dimension to match the input image size. Let S represent the k-th saliency map of the source image, where ReLU is the activation function and S represents the source image. Since this paper is only interested in features that positively influence the category to be identified, the ReLU activation function is applied to Grad-CAM saliency map generation, selecting only features that increase the category score y. c Positive pixels. Without ReLU, Grad-CAM saliency maps sometimes highlight not only the desired categories, but also perform worse in target localization.

[0065] Step 2: Based on the Grad-CAM saliency map of the source image, A mask is generated from the source image. To cut out a selected rectangular region from the source image and paste it into the corresponding position in the target image, a new region is created on the target image. A binary mask matrix of the same size as the target image needs to be generated, where the elements corresponding to the cut and pasted regions have a value of 1, and the elements of other regions have a value of 0. The mask matrix is ​​then multiplied element-wise with the target image. This means that only the pixel values ​​corresponding to the cut and pasted regions will be retained in the target image, while the pixel values ​​of other regions will be adjusted according to the mask matrix. The specific steps are as follows: A random variable λ following a Beta distribution is randomly selected to control the degree of blending. The larger λ is, the smaller the cut region is, and the larger the proportion of the source image; the smaller λ is, the larger the cut region is, and the larger the proportion of the target image. The width r of the cut region is calculated based on λ and the width W and height H of the image. w and height r h , making r w r h WH = 1 - λ. The coordinates r of the center point of the randomly selected shearing region are... x r y ), making r x ~Unif(0, W), r y ~Unif(0, H). Based on the center point coordinates, width, and height, determine the bounding box B of the clipping region = (r... x r y ,,r w rh Initialize the mask matrix M as an all-1 matrix, then set the elements in M ​​that are inside B to 0, and keep the elements outside B to 1. This completes the sampling of the mask matrix M.

[0066] Step 3: Extract the foreground information of the source image as a patch based on the mask of the source image. Assume I s ∈R W×H×C The source image is randomly selected, where W, H, and C represent the image width, height, and number of channels, respectively, and the label is y. s The sampling blending ratio λ is generated from the Beta(α, α) distribution, which determines the percentage of the image to be cropped. The ultimate goal is to cut a patch of size I from the source image. For the source image I... s The saliency detection can be represented as:

[0067]

[0068] in, Representative source image I s The saliency plot, f Grad-CAM (·) represents the Grad-CAM saliency map generation model. Then, the pixel with the maximum saliency value is searched within the saliency map. M represents the mask matrix, which can be found using the following formula:

[0069]

[0070] Here, argmax(·) represents the parameter function that takes the maximum value. If possible, select the source image with... A patch centered on the element; otherwise, select include. This is a patch. This ensures that the patch is selected from the target object area, i.e., the foreground, and not from the background.

[0071] Step 4: Input the target image and generate a saliency map using the lightweight method Grad-CAM. This is to obtain the Grad-CAM saliency map for category c of the target image. This paper first calculates the score gradient y for category c. c and the feature mapping A of the last convolutional layer k Correspondence

[0072]

[0073] The gradients from these backpropagations are then calculated using global average pooling across the width and height dimensions (indexed by i and j, respectively) to obtain the weights of the k-th feature map for class c.

[0074]

[0075] Where Z is the number of pixels in the feature map. y represents the weight of the k-th feature map. c Let A represent the score gradient of category c. k This is the feature map of the k-th channel after the last feature mapping layer of the convolutional layer. A represents k The (i,j)th feature point is obtained. After calculating the weights of the class on all feature maps, their weighted sums are applied, followed by a ReLU activation function and upsampling to obtain the final Grad-CAM saliency map.

[0076]

[0077] Where φ(·) represents the operation of upsampling along the feature map dimension to match the input image size. Let represent the k-th saliency map of the target image, and T represent the target image. Since this paper is only interested in features that positively influence the category to be identified, the ReLU activation function is applied to Grad-CAM saliency map generation, selecting only features that increase the category score y. c Positive pixels. Without ReLU, Grad-CAM saliency maps sometimes highlight not only the desired categories, but also perform worse in target localization.

[0078] Example 2, Step 5: Use the foreground information extracted from the source image as a patch for composite blending in the Grad-CAM region of the target image. The source image is the image to be cut and pasted, while the target image is the portion of the source image that receives the data. Assume I t ∈R W×H×C It is another randomly selected target image, labeled y. t This is used to blend with the source image patch. Blend I t and I s To generate new training samples I mix This refers to image blending. The blending of two images can be defined as:

[0079] I mix =M⊙I s +(1-M)⊙I t (18)

[0080] Among them, I mix Denotes a blended image, M∈{0,1} W×H⊙ represents a binary mask, and ⊙ represents element-wise multiplication. Since the patch positions of the source image λ×W×H have already been calculated, the corresponding positions in mask M are set to 1, and the other positions are set to 0. The element-wise multiplication of M with the source image results in an image with all content except the source image patch positions removed. (1-M) is the opposite of M; that is, the element-wise multiplication of (1-M) with the target image preserves all regions except the source image patch positions. Finally, adding the two together creates a new blended image sample containing the target image with the selected source patch, resulting in the final enhanced image.

[0081] Step Six: Calculate the new label for the final enhanced image. Specifically, this involves normalizing the Grad-CAM saliency map. Using the positions of the source image patches mentioned earlier, calculate the semantic percentage of each patch position in the original image relative to the source and target images. Output these semantic percentages as the new label for the final enhanced image. The generated Grad-CAM saliency map is normalized, and the normalized saliency map is defined as a semantic information measurement map, i.e., a semantic percentage map (SPM). SPM quantifies the percentage of correlation between pixels and labels. This section calculates the image's SPM in the following way:

[0082]

[0083] in This represents the Grad-CAM saliency map for category c.

[0084] For enhanced image I mix The labels are calculated in this paper at the locations of generated source image patches in both the source and target images. The proportions of source and target images in the enhanced image labels are as follows:

[0085]

[0086]

[0087] Where M is the source image patch location mask in the enhanced image generation process. To increase the proportion of source images in image tags, To increase the proportion of the target image in the image label, the corresponding mask for the target image is (1-M). Finally, the label for the enhanced image is:

[0088]

[0089] Among them, y mixTo enhance image labeling, the process described above generates label information that better reflects the image's inherent semantic composition. In fine-grained recognition, although the discriminative information of an image is highly non-uniformly distributed spatially, the method described in this section prevents the introduction of significant label noise into the enhanced data. Furthermore, the two components of an enhanced image typically do not converge in semantic proportion. Therefore, unlike CutMix, this method does not restrict label coefficients. and The sum is 1.

[0090] Figure 3 This diagram illustrates the network performance of traditional image augmentation using different training strategies. In deep learning, loss is a metric used to measure the difference between the model's predictions and the actual labels. It represents the degree of error or inaccuracy of the model's predictions during training. By minimizing the loss function, we can optimize the model's parameters to more accurately predict the target variable. To verify the effectiveness of each module in the saliency map-guided fine-grained data augmentation method, this invention uses ResNet50 as the base network and adds different modules and strategies to the CutMix method, testing it on the CUB dataset. The experimental results are as follows: Figure 3 As shown, although the lightweight Grad-CAM saliency map guidance does not offer a significant advantage over conventional saliency map guidance, it offers a clear benefit compared to CutMix alone. Semantic assignment of labels significantly enhances the CutMix data augmentation method, and combining Grad-CAM saliency map guidance with semantic label assignment can achieve substantial improvements.

[0091] Furthermore, in order to provide a concise description of exemplary embodiments, not all features of actual embodiments (i.e., those features that are not relevant to the currently considered best mode for carrying out the invention, or those features that are not relevant to implementing the invention) may be omitted.

[0092] It should be understood that numerous specific implementation decisions can be made during the development of any practical implementation, such as in any engineering or design project. Such development efforts may be complex and time-consuming, but for those skilled in the art who benefit from this disclosure, the development effort will be a routine task in design, manufacturing, and production without requiring extensive experimentation.

[0093] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A fine-grained data augmentation method based on saliency map guidance, characterized in that: This method is applied to fine-grained image classification; it selects the foreground of the source image for blending based on saliency maps to avoid introducing invalid regions; and it utilizes Grad-CAM semantic information to evaluate the labels of the newly generated blended images, reducing label noise. The method includes the following steps: Step 1: Input the source image and generate a saliency map using the lightweight method Grad-CAM; Step 2: Generate a mask for the source image based on the Grad-CAM saliency map of the source image; Step 3: Extract the foreground information of the source image based on the mask of the source image; Step 4: Input the target image and generate a saliency map using the lightweight method Grad-CAM; Step 5: Use the foreground information extracted from the source image as a patch to synthesize the Grad-CAM region of the target image; Step Six: Calculate the new labels for the final enhanced image; The method locates foreground regions with speech meaning in the source and target images, and then performs data augmentation by replacing significant semantic information in the source and target images. The designed data augmentation method can locate foreground information to achieve diversified data expansion, thereby improving the robustness of the data augmentation method. Step one, which uses the lightweight Grad-CAM method to generate a saliency map, can be represented in the following form: (1) Where Z is the number of pixels in the feature map. y represents the weight of the k-th feature map. c This represents the score gradient for category c. This is the feature map of the k-th channel after the last feature mapping layer of the convolutional layer. express The (i,j)th feature point in the middle; Weights of class c based on the k-th feature map The Grad-CAM saliency map is shown below: (2) in, (·) indicates an upsampling operation along the feature map dimension to match the input image size. Represents the saliency map of the source image, where ReLU is the activation function; Step two is based on the saliency map of the source image. Generate a mask; perform thresholding on the saliency map, select an appropriate threshold based on the intensity of pixels in the saliency map, and generate a binary mask; Step three involves extracting the foreground information of the source image based on the mask of the source image; The saliency detection of the source image Is can be represented as: (3) in, The saliency map representing the source image Is, f Grad-CAM (·) represents the Grad-CAM saliency map generation model; it searches for the pixel with the largest saliency value in the saliency map; M represents the mask matrix. One way to represent the dimension of an image can be found using the following formula: (4) in, This represents the function that takes parameters; Step four involves inputting the target image and generating a saliency map using the lightweight Grad-CAM method: (5) in, (·) indicates an upsampling operation along the feature map dimension to match the input image size. This represents the weight of the k-th feature map. Let A represent the k-th saliency map of the target image. k This is the feature mapping for the last convolutional layer, where T represents the target image; In step five, the foreground information extracted from the source image is used as a patch for synthesis in the Grad-CAM region of the target image; The blending of two images can be defined as: (6) Among them, I mix Represents a mixed image. This represents a binary mask, and ⊙ represents element-wise multiplication. It is the target image; In step six, the new label for the final enhanced image is calculated; First, the generated Grad-CAM saliency map is normalized, and the percentage map SPM of the image is calculated as follows: (7) in This represents the Grad-CAM saliency map for category c; The proportions of source and target images in the enhanced image labels are as follows: (8) Where M is the source image patch location mask in the enhanced image generation process. To increase the proportion of source images in image tags, To increase the proportion of the target image in the image labels, the final labels for the enhanced images are: (9) in, To enhance image labeling, These are the labels for the target image; Through the above process, the label information generated for enhanced images can better reflect their inherent semantic composition.

Images