Multi-label image classification method and device, readable storage medium and electronic equipment

Abstract

The invention discloses a multi-label image classification method and device, a readable storage medium and electronic equipment, and the method comprises the steps: inputting preprocessed image data into a pre-trained convolutional neural network model, and extracting a feature map; performing convolution operation on the extracted feature map to obtain a category activation map; determining a bounding box of a target object in the category activation graph, and extracting a local region image from the feature graph according to coordinates of the bounding box; adjusting the local area image to a preset size by adopting a bilinear interpolation method; screening out a local area image smaller than a threshold size, and adjusting the screened image to a preset size by adopting a bicubic interpolation method; and inputting the local region image after size adjustment into the convolutional neural network model to extract a feature map again, and performing image classification prediction according to the feature map extracted again. According to the invention, richer feature information can be obtained, so that the model can identify the small target object in the image more accurately.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a multi-label image classification method, device, readable storage medium and electronic device. Background Art

[0002] In the field of computer vision, image classification technology has always been a hot topic of research. Traditional image classification tasks focus on identifying a single category in an image. This classification method assumes that each image has only one category and is classified into a mutually exclusive category label. However, images in the real world often contain objects of multiple categories, which may belong to different categories, making the multi-label image classification problem complex. In the multi-label image classification task, the model needs to predict the categories of all objects present in the image, and the output result is a set of non-mutually exclusive category attributes.

[0003] Due to the diversity and complexity of objects in real images, it is often difficult to capture all objects by relying solely on global feature extraction. At the same time, there are even more difficult to capture small target areas in the image. Therefore, this method of extracting only global features is prone to model omission problems when performing image classification. Summary of the invention

[0004] In view of the above situation, in order to improve the accurate recognition of small targets, the present invention proposes a multi-label image classification method, device, readable storage medium and electronic device.

[0005] The present invention provides a multi-label image classification method, comprising: Input the preprocessed image data into the pre-trained convolutional neural network model and extract the feature map; Performing a convolution operation on the extracted feature map to obtain a category activation map; Determine a bounding box of the target object in the class activation map, and extract a local area image from the feature map according to the coordinates of the bounding box; Adopting bilinear interpolation method to adjust the local area image to a preset size; Screening out local area images smaller than a threshold size from the extracted local area images, and adjusting the screened images to a preset size using a bicubic interpolation method; The resized local area image is input into the convolutional neural network model to extract feature maps again, and image classification prediction is performed based on the re-extracted feature maps.

[0006] Furthermore, in the above multi-label image classification method, the step of determining the bounding box of the target object in the class activation map comprises: Normalizing the score vector of each category region in the category activation map; By comparing the normalized score vector with the set threshold, the bounding box of the target object is determined.

[0007] Furthermore, in the above multi-label image classification method, the score vector of each category area in the category activation map includes a horizontal score vector and a vertical score vector, and the horizontal score vector is: ; The vertical score vector is: , Where F is the category activation map, i is the i-th category area in the category activation map, x is the horizontal coordinate in the category activation map, y is the vertical coordinate in the category activation map, h1 and h2 are the horizontal and vertical lengths of the category activation map, respectively.

[0008] Furthermore, in the above multi-label image classification method, after the step of extracting the local area image from the feature map according to the coordinates of the bounding box, the step further includes: Determine the center position coordinates of each of the local area images, and calculate the Euclidean distance between the center positions of each of the local area images; Establish the spatial relationship between the objects based on the Euclidean distance between the center positions of all local area images; The step of performing image classification prediction based on the re-extracted feature map comprises: The re-extracted feature map is fused with the spatial relationship between the target object, and the prediction result of image classification is determined based on the fusion result.

[0009] The present invention also discloses a multi-label image classification device, comprising: A feature extraction module is used to input the preprocessed image data into the pre-trained convolutional neural network model and extract feature maps; A convolution module, used for performing a convolution operation on the extracted feature map to obtain a category activation map; A local area image extraction module, used to determine a bounding box of a target object in the class activation map, and extract a local area image from the feature map according to the coordinates of the bounding box; A first image adjustment module, which uses a bilinear interpolation method to adjust the local area image to a preset size; A second image adjustment module, used for screening out local area images with a size smaller than a threshold from the extracted local area images, and adjusting the screened images to a preset size by using a bicubic interpolation method; The image classification module is used to input the local area image after resizing into the convolutional neural network model to extract the feature map again, and perform image classification prediction based on the re-extracted feature map.

[0010] Furthermore, in the above-mentioned multi-label image classification device, the step of determining the bounding box of the target object in the class activation map includes: Normalizing the score vector of each category region in the category activation map; By comparing the normalized score vector with the set threshold, the bounding box of the target object is determined.

[0011] Furthermore, the multi-label image classification device further includes: A calculation module, used to determine the center position coordinates of each of the local area images, and calculate the Euclidean distance between the center positions of each of the local area images; A spatial relationship establishment module, used to establish the spatial relationship between the targets according to the Euclidean distance between the center positions of all local area images; The step of performing image classification prediction based on the re-extracted feature map comprises: The re-extracted feature map is fused with the spatial relationship between the target object, and the prediction result of image classification is determined based on the fusion result.

[0012] The present invention also discloses a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the multi-label image classification method described in any one of the above items is implemented.

[0013] The present invention also discloses an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the multi-label image classification methods described above when executing the computer program.

[0014] The present invention upsamples the small target area through bilinear and bicubic interpolation methods to obtain richer feature information, and can more accurately identify small target objects in the image. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flowchart of a multi-label image classification method provided by the first embodiment of the present invention; Figure 2 A flowchart of a multi-label image classification method provided by a second embodiment of the present invention; Figure 3 A structural block diagram of a multi-label image classification device provided by a third embodiment of the present invention; Figure 4FIG. 4 is a schematic diagram of the structure of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0017] With reference to the following description and accompanying drawings, embodiments of the present invention will be clear. In these descriptions and accompanying drawings, some specific implementations of the embodiments of the present invention are specifically disclosed to represent some ways of implementing the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto. On the contrary, embodiments of the present invention include all changes, modifications and equivalents that fall within the spirit and connotation of the appended claims.

[0018] See also Figure 1 , is a multi-label image classification method in the first embodiment of the present invention, comprising steps S11 to S16.

[0019] Step S11, input the preprocessed image data into the pre-trained convolutional neural network model and extract the feature map.

[0020] The image data is the image to be detected. The image data is preprocessed, mainly to adjust the size of all input images to a uniform specification, for example, to a size of 448*448*3 pixels. This embodiment can use a commonly used convolutional neural network model to extract image features, for example, a ResNext-101 model can be used. The ResNext-101 model is pre-trained on the ImageNet dataset. There are four intermediate layers in ResNext-101, layer1, layer2, layer3, and layer4. Extract feature maps from the last layer, layer4 , where w, h, and c are the width, height, and number of channels of the feature map, respectively. Specifically, the feature map is extracted in layer4 , the pixel size of the feature map is 14*14*2048.

[0021] Step S12: performing a convolution operation on the extracted feature map to obtain a category activation map.

[0022] The feature map is directly subjected to a 1*1 convolution (i.e., 1*1Conv) operation to obtain a class activation map, the size of which is, for example, 14*14 pixels. The class activation map is used to locate areas in the image that are related to a specific category, and to provide target location information for the local area and small target area in the next step.

[0023] Step S13, determining a bounding box of the target object in the class activation map, and extracting a local area image from the feature map according to the coordinates of the bounding box.

[0024] Specifically, in one implementation of the present invention, the step of determining a bounding box of a target object in the class activation map includes: Normalizing the score vector of each category region in the category activation map; By comparing the normalized score vector with the set threshold, the bounding box of the target object is determined.

[0025] The score vector includes a horizontal score vector and a vertical score vector. The horizontal score vector of each category area in the category activation map and the vertical score vector , expressed as: , , F is the category activation map, i is the i-th category area in the category activation map, x is the horizontal coordinate in the category activation map, y is the vertical coordinate in the category activation map, h1 and h2 are the horizontal and vertical lengths of the category activation map, respectively. and It represents the position of each category area in the activation map, taking the maximum activation value at that position.

[0026] And, the score vector is normalized to the range of 0 to 1. The normalization formula is: and ,in, and are the minimum and maximum eigenvalues ​​on the horizontal axis, and are the minimum and maximum eigenvalues ​​on the ordinate, respectively.

[0027] By setting a threshold τ, we determine the areas where the score vector exceeds the threshold and calculate the and as well as and To obtain the local area image, we need to analyze the score vector of the category area and the threshold value. The relationship between τ and τ is used to distinguish local areas, where the part greater than the threshold is positive and the part less than or equal to the threshold is negative. In the category activation map, the higher the horizontal and vertical scores of the region, the more likely it is to be part of the target. By setting a threshold τ, the area with a score higher than the threshold is considered a positive area, and the area with a score lower than the threshold is considered a negative area. Therefore, the position where "positive turns negative" refers to the position in the score vector that suddenly changes from above the threshold (positive area) to below the threshold (negative area). According to the distribution of the position where the positive turns negative, the bounding box of the target object is selected.

[0028] It should be noted that if there is no position where the positive value changes to the negative value, a part of the entire score vector is used as the bounding box of the target by default. That is to say, although the score vector has no obvious dividing point, the approximate position of the target object is still determined based on the area with a high score in the activation map, and the bounding box of this area is obtained.

[0029] Step S14: adjusting the local area image to a preset size using a bilinear interpolation method.

[0030] Step S15, screening out local area images with a size smaller than a threshold from the extracted local area images, and adjusting the screened images to a preset size using a bicubic interpolation method.

[0031] In specific implementation, the size of the screened local area image can be adjusted to an image of 448*448*3 pixels by bilinear interpolation. At the same time, special considerations are also made for smaller targets in the image, that is, a size standard is defined to identify small targets. This standard is, for example, 10% of the original image size, that is, if the length and width of the target are both less than 10% of the original image size, then the target is classified as a small target. Once the small target is identified, the regional images of these small targets are resized using the bicubic interpolation method to make them 448*448*3 pixel images. That is, in this embodiment, each target area is restored by two different upsampling methods, which is conducive to obtaining more information about the small target area.

[0032] Step S16, inputting the resized local area image into the convolutional neural network model to extract the feature map again, and performing image classification prediction based on the re-extracted feature map.

[0033] The resized local area image is input into the convolutional neural network model, and the image features are extracted through the convolutional neural network model. That is, the local area image is firstly extracted with the ResNext-101 model to obtain the feature map, and the obtained feature map is subjected to pooling, convolution and Softmax functions to obtain the prediction result.

[0034] This embodiment upsamples the small target area through bilinear and bicubic interpolation methods to obtain richer feature information, which can more accurately identify small target objects in the image.

[0035] Furthermore, the spatial relationship between objects in an image plays a significant role in the accuracy of image classification and recognition. Traditional image classification methods often ignore these subtle but critical spatial relationships, thereby limiting the performance of model recognition. Therefore, in order to make full use of the positional relationship between objects and improve the recognition accuracy of small objects, a multi-label image classification method is proposed in the second embodiment of the present invention, such as Figure 2 As shown, steps S21 to S28 are included.

[0036] Step S21, input the preprocessed image data into the pre-trained convolutional neural network model and extract the feature map.

[0037] Step S22: performing a convolution operation on the extracted feature map to obtain a category activation map.

[0038] Step S23, determining a bounding box of the target object in the class activation map, and extracting a local area image from the feature map according to the coordinates of the bounding box.

[0039] Step S24: adjusting the local area image to a preset size using a bilinear interpolation method.

[0040] Step S25, screening out local area images with a size smaller than a threshold from the extracted local area images, and adjusting the screened images to a preset size using a bicubic interpolation method.

[0041] It can be understood that the implementation method of steps S21 to S25 can refer to the implementation method in the first embodiment and will not be described in detail here.

[0042] Step S26, determining the center position coordinates of each of the local area images, and calculating the Euclidean distance between the center positions of each of the local area images.

[0043] Step S27: establishing the spatial relationship between the targets according to the Euclidean distances between the center positions of all the local area images.

[0044] Step S28, input the resized local area image into the convolutional neural network model to extract the feature map again, and fuse the re-extracted feature map with the spatial relationship between the target object, and determine the prediction result of the image classification according to the fusion result.

[0045] Specifically, the coordinates of the center position of each local area image obtained above are used to determine the coordinates of its center position, and the Euclidean distance is used to learn the spatial relationship between different targets. Specifically, the center position coordinates of each local area image obtained are calculated by the following formula and , the formula is as follows: ; , and is the minimum and maximum value of the x-axis on the local area image, and is the minimum and maximum value of the y-axis on the local area image.

[0046] For each local region image, calculate the Euclidean distance between it and the center position of another local region image , the calculation formula is as follows: , ( , ), , ) represent the center position coordinates of the ith local area image and the jth local area image respectively.

[0047] Furthermore, the obtained center position relationship is normalized.

[0048] , in, is the minimum Euclidean distance between locations, is the maximum Euclidean distance between locations.

[0049] When the Euclidean distance is smaller, it means that the correlation between the two targets is stronger, and the obtained A value is larger. The expression is as follows: .

[0050] Calculate the Euclidean distance between the center positions of all local area images and form a category area correlation matrix , k represents the category correlation matrix between image categories in a batch of images.

[0051] The resized local area image is input into the ResNext-101 model (convolutional neural network model), the feature map is extracted through the ResNext-101 model, the extracted feature map is fused with the category area correlation matrix, and the fusion result is used to obtain the prediction result through convolution operation and Softmax function.

[0052] Specifically, the category-area correlation matrix The feature maps extracted by the ResNext-101 model are used as input and applied to the graph convolutional network (GCN) to obtain the l+1 layer output, i.e., z l+1 , where z l+1 Each element in belongs to the interval [0,1]. After GCN processing, the image feature relationship matrix containing all category position relationships is obtained, as follows: , Among them, Z l , U l and f l They represent the input feature matrix of the first layer in GCN, the weight matrix of the lth layer, and the nonlinear activation function of the lth layer of GCN respectively.

[0053] The image feature relationship matrix is ​​subjected to convolution operation and Softmax function to obtain the prediction result.

[0054] The beneficial effects of this embodiment are specifically described as follows: (1) The small target area is upsampled through bilinear and bicubic interpolation methods to obtain richer feature information, so that the model can more accurately identify small target objects in the image.

[0055] (2) Using Euclidean distance to learn the spatial relationship between objects enhances the model’s understanding of the positional relationship of objects in the image, which helps improve classification accuracy.

[0056] See also Figure 3 , is a multi-label image classification device in a third embodiment of the present invention, comprising: A feature extraction module 31 is used to input the preprocessed image data into a pretrained convolutional neural network model and extract a feature map; A convolution module 32, configured to perform a convolution operation on the extracted feature map to obtain a class activation map; A local area image extraction module 33, used to determine a bounding box of the target object in the class activation map, and extract a local area image from the feature map according to the coordinates of the bounding box; A first image adjustment module 34, which uses a bilinear interpolation method to adjust the local area image to a preset size; A second image adjustment module 35 is used to filter out local area images with a size smaller than a threshold from the extracted local area images, and adjust the filtered images to a preset size using a bicubic interpolation method; The image classification module 36 is used to input the local area image after resizing into the convolutional neural network model to extract the feature map again, and perform image classification prediction based on the re-extracted feature map.

[0057] Furthermore, in the above-mentioned multi-label image classification device, the step of determining the bounding box of the target object in the class activation map includes: Normalizing the score vector of each category region in the category activation map; By comparing the normalized score vector with the set threshold, the bounding box of the target object is determined.

[0058] Furthermore, the multi-label image classification device further includes: A calculation module, used to determine the center position coordinates of each of the local area images, and calculate the Euclidean distance between the center positions of each of the local area images; A spatial relationship establishment module, used to establish the spatial relationship between the targets according to the Euclidean distance between the center positions of all local area images; The step of performing image classification prediction based on the re-extracted feature map comprises: The re-extracted feature map is fused with the spatial relationship between the target object, and the prediction result of image classification is determined based on the fusion result.

[0059] The multi-label image classification device provided in the embodiment of the present invention has the same implementation principle and technical effects as those in the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the device embodiment, reference can be made to the corresponding contents in the aforementioned method embodiment.

[0060] Another aspect of the present invention provides an electronic device, see Figure 4 , shown is an electronic device in an embodiment of the present invention, including a processor 10, a memory 20, and a computer program 30 stored in the memory and executable on the processor, wherein when the processor 10 executes the computer program 30, the multi-label image classification method as described above is implemented.

[0061] The electronic device may be, but is not limited to, a computer device such as a personal computer, a mobile phone, etc. In some embodiments, the processor 10 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip, for running program codes stored in the memory 20 or processing data.

[0062] Among them, the memory 20 includes at least one type of readable storage medium, and the readable storage medium includes a flash memory, a hard disk, a multimedia card, a card-type memory (for example, an SD or DX memory, etc.), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory 20 can be an internal storage unit of an electronic device, such as a hard disk of the electronic device. In other embodiments, the memory 20 can also be an external storage device of an electronic device, such as a plug-in hard disk equipped on the electronic device, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), etc. Further, the memory 20 can also include both an internal storage unit of the electronic device and an external storage device. The memory 20 can not only be used to store application software and various types of data installed in the electronic device, but also can be used to temporarily store data that has been output or is to be output.

[0063] Optionally, the electronic device may also include a user interface, a network interface, a communication bus, etc. The user interface may include a display (Display), an input unit such as a keyboard (Keyboard), and the optional user interface may also include a standard wired interface and a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, and an OLED (Organic Light-Emitting Diode) touch device, etc. Among them, the display may also be appropriately referred to as a display screen or a display unit, which is used to display information processed in the electronic device and to display a visual user interface. The network interface may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface), which is usually used to establish a communication connection between the device and other electronic devices. The communication bus is used to realize the connection and communication between these components.

[0064] It should be pointed out that Figure 4 The structure shown does not constitute a limitation on the electronic device. In other embodiments, the electronic device may include fewer or more components than those shown in the figure, or combine certain components, or arrange the components differently.

[0065] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which implements the multi-label image classification method as described above when the program is executed by a processor. Those skilled in the art will appreciate that the logic and / or steps represented in the flowchart or otherwise described herein, for example, may be considered as an ordered list of executable instructions for implementing logical functions, and may be specifically implemented in any computer-readable medium for use by an instruction execution system or device (such as a computer-based system, a system including a processor, or other system that can obtain instructions from an instruction execution system or device and execute instructions), or in combination with such instruction execution systems or devices. For purposes of this specification, "computer-readable medium" may be any device that can contain, store, communicate, propagate, or transmit a program for use by an instruction execution system or device, or in combination with such instruction execution systems or devices.

[0066] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0067] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or a combination thereof: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0068] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0069] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A multi-label image classification method, characterized in that: include, Input the preprocessed image data into the pre-trained convolutional neural network model and extract the feature map; Performing a convolution operation on the extracted feature map to obtain a category activation map; Determine a bounding box of the target object in the class activation map, and extract a local area image from the feature map according to the coordinates of the bounding box; Adopting bilinear interpolation method to adjust the local area image to a preset size; Screening out local area images smaller than a threshold size from the extracted local area images, and adjusting the screened images to a preset size using a bicubic interpolation method; The resized local area image is input into the convolutional neural network model to extract feature maps again, and image classification prediction is performed based on the re-extracted feature maps.

2. The multi-label image classification method according to claim 1, characterized in that: The step of determining the bounding box of the target object in the class activation map comprises: Normalizing the score vector of each category region in the category activation map; By comparing the normalized score vector with the set threshold, the bounding box of the target object is determined.

3. The multi-label image classification method according to claim 2, characterized in that: The score vector of each category region in the category activation map includes a horizontal score vector and a vertical score vector, and the horizontal score vector is: ； The vertical score vector is: ， Where F is the category activation map, i is the i-th category area in the category activation map, x is the horizontal coordinate in the category activation map, y is the vertical coordinate in the category activation map, h1 and h2 are the horizontal and vertical lengths of the category activation map, respectively.

4. The multi-label image classification method according to claim 1, characterized in that: After the step of extracting the local area image from the feature map according to the coordinates of the bounding box, the method further includes: Determine the center position coordinates of each of the local area images, and calculate the Euclidean distance between the center positions of each of the local area images; Establish the spatial relationship between the objects based on the Euclidean distance between the center positions of all local area images; The step of performing image classification prediction based on the re-extracted feature map comprises: The re-extracted feature map is fused with the spatial relationship between the target object, and the prediction result of image classification is determined based on the fusion result.

5. A multi-label image classification device, characterized in that: include: A feature extraction module is used to input the preprocessed image data into the pre-trained convolutional neural network model and extract feature maps; A convolution module, used for performing a convolution operation on the extracted feature map to obtain a category activation map; A local area image extraction module, used to determine a bounding box of a target object in the class activation map, and extract a local area image from the feature map according to the coordinates of the bounding box; A first image adjustment module, which uses a bilinear interpolation method to adjust the local area image to a preset size; A second image adjustment module, used for screening out local area images with a size smaller than a threshold from the extracted local area images, and adjusting the screened images to a preset size by using a bicubic interpolation method; The image classification module is used to input the local area image after resizing into the convolutional neural network model to extract the feature map again, and perform image classification prediction based on the re-extracted feature map.

6. The multi-label image classification device according to claim 5, characterized in that: The step of determining the bounding box of the target object in the class activation map comprises: Normalizing the score vector of each category region in the category activation map; By comparing the normalized score vector with the set threshold, the bounding box of the target object is determined.

7. The multi-label image classification device according to claim 5, characterized in that: Also includes: A calculation module, used to determine the center position coordinates of each of the local area images, and calculate the Euclidean distance between the center positions of each of the local area images; A spatial relationship establishment module, used to establish the spatial relationship between the targets according to the Euclidean distance between the center positions of all local area images; The step of performing image classification prediction based on the re-extracted feature map comprises: The re-extracted feature map is fused with the spatial relationship between the target object, and the prediction result of image classification is determined based on the fusion result.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the multi-label image classification method according to any one of claims 1 to 4 is implemented.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the multi-label image classification method according to any one of claims 1 to 4 is implemented.

Images