An adaptive coordinated attention network-based method for fatty liver ultrasound image classification

Abstract

The present invention discloses a method for classifying fatty liver B-ultrasound images using an adaptive coordinated attention network, including a base plate. The method is designed to perform data enhancement preprocessing operations such as image histogram equalization, grayscale linear transformation, ROI selection, and average grayscale ratio comparison. The method uses an EfficientNet network combined with label smoothing and Adaptive Coordinate Attention to only use the pixels and labels of the original B-ultrasound image as input, thereby avoiding the decoupling of feature extraction and classification and achieving end-to-end automatic classification and recognition. The method is a lightweight network with fast training and reasoning, small model parameters, and based on adaptive coordinated attention, the method can adaptively encode horizontal and vertical position information as channel attention, while retaining position information in other spatial directions to enhance the representation of the attention object. Based on the method, the recognition ability of fatty liver B-ultrasound images has exceeded manual recognition and is also superior to other methods.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a fatty liver B-ultrasound image classification method based on an adaptive coordinated attention network. Background Art

[0002] Medical analysis of B-ultrasound images of fatty liver is a key research tool for the diagnosis and analysis of fatty liver, and is also one of the important challenges for the medical classification of fatty liver. For a long time, many researchers have been identifying various normal and abnormal manifestations of fatty liver and other related research. However, the current analysis of fatty liver images to determine the degree of fatty liver mainly relies on the manual judgment of doctors, which is subjective, time-consuming, and inevitably leads to misdiagnosis. Therefore, the efficiency of manual judgment of the degree of fatty liver is not high.

[0003] Currently, most of the existing classification methods for fatty liver B-ultrasound images are based on traditional machine learning methods. Traditional methods require manual design or feature extraction. Extracting features of fatty liver B-ultrasound images takes a long time and involves complex steps. Feature extraction and classification recognition are decoupled. In addition, the classification recognition results obtained by the features obtained by this method have low accuracy, and the obtained model has poor generalization performance and poor robustness. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a fatty liver B-ultrasound image classification method based on an adaptive coordinated attention network to facilitate the cleaning of river silt.

[0005] The present invention adopts the following technical solutions to achieve the invention objectives:

[0006] A method for classifying fatty liver B-ultrasound images using an adaptive coordinated attention network, characterized in that it comprises the following steps:

[0007] Step (1) preprocessing the obtained fatty liver B-ultrasound image to obtain a data enhanced B-ultrasound image;

[0008] Step (2) extracting the Region Of Interest of the fatty liver B-ultrasound image, performing far-field and near-field segmentation, and performing far-field and near-field grayscale analysis;

[0009] Step (3) performing data segmentation on the preprocessed B-ultrasound image;

[0010] Step (4) Design the Adaptive Coordinate Attention structure, use label smoothing, rebuild the EfficientNet model, construct MBCABlock to improve the network modeling capability, feed the preprocessed training set images into the network for training, and adjust the network structure and hyperparameters through the validation set;

[0011] Step (5) uses the trained model to predict images that the model has never seen, namely the test set data, to test the generalization ability of the method.

[0012] As a further limitation of the present technical solution, in step (1), a series of preprocessing operations are performed on the fatty liver B-ultrasound image, using adaptive grayscale histogram equalization, grayscale linear transformation, and Gaussian smoothing noise reduction, which are specifically as follows:

[0013] Step (1.1) adjusts the image size from an irregular size to 300*300, and at the same time improves the image resolution;

[0014] Step (1.2) grayscale processing is performed on the image, grayscale distribution is counted, and an algorithm is used to adaptively equalize grayscale histograms, so as to improve near-field and far-field contrast in image display;

[0015] r and s represent the grayscale of the original image and the grayscale of the image after histogram equalization, respectively.

[0016] The formula for adaptive grayscale histogram equalization is as follows:

[0017] s=T(r) r=T -1 (s) (1)

[0018]

[0019]

[0020]

[0021]

[0022] Step (1.3) performs a grayscale linear transformation on the image after histogram equalization. Since the original image has noise blur and various impurity noises, this method is used to highlight the required features while reducing irrelevant features.

[0023] The grayscale linear transformation formula in step (1.3) is as follows:

[0024]

[0025] Where: f(x, y) is the gray value of the input image;

[0026] g(x, y) is the gray value of the output image;

[0027] [a, b] is the grayscale value range of the input image;

[0028] [c, d] is the grayscale value range of the output image;

[0029] By adjusting the intercept of the function on the Y axis, the grayscale values ​​of all pixels are moved up or down, so as to make the entire image darker or brighter.

[0030] Step (1.4) performs Gaussian smoothing noise reduction on the processed image. Gaussian filtering is a linear smoothing filter that has a good suppressive effect on noise that obeys a normal distribution.

[0031] As a further limitation of the present technical solution, the step (2) specifically includes the following operations:

[0032] Step (2.1) performs near-field and far-field segmentation on the processed image, segmenting the fatty liver B-ultrasound image into near field, mid-field and far field, and extracting two ROIs from each B-ultrasound image, one in the near field and the other in the far field;

[0033] Step (2.2) calculates the average grayscale of the extracted ROI, which can reflect many details of the image;

[0034] The average grayscale calculation formula in step (2.2) is as follows:

[0035]

[0036]

[0037] Where: n k is the grayscale z k The number of times it appears in the image;

[0038] MN is the total number of pixels;

[0039] Step (2.3) performs MIR feature calculation on the calculated ROI average grayscale of different field areas and calculates the near-field and far-field average grayscale ratio. According to the characteristics of near-field echo enhancement and far-field echo attenuation of B-ultrasound images of fatty liver, the selection of MIR features can well reflect the severity of fatty liver.

[0040] The MIR feature calculation formula in step (2.3) is as follows:

[0041]

[0042]

[0043] Where: H and W are the height and width of ROI respectively;

[0044] In this paper, it is a fixed value of 100, x and y represent the coordinates of the pixel point, and I(x, y) represents the grayscale value of the pixel point in the ROI.

[0045] As a further limitation of the present technical solution, the specific operation of step (3) is as follows:

[0046] 70% of the data set of each type of image is randomly selected as the training set, 20% of the images are used as the validation set, and the remaining 10% are used as the test set. The test set is used to evaluate the model's ability to discriminate new samples, and the test error is used as an approximation of the generalization error.

[0047] As a further limitation of the technical solution, the specific operation of step (4) is as follows:

[0048] Step (4.1) Design the Adaptive Coordinate Attention structure to adaptively encode the horizontal and vertical position information of the image into the channel attention. While capturing long-range dependencies along one spatial direction, it can also retain precise position information along the other spatial direction. Adaptive learning is performed on the information in the two spatial directions to obtain different degrees of weight information.

[0049] The specific implementation formula of step (4.1) Adaptive Coordinate Attention is as follows:

[0050]

[0051] f=δ(F1([z h , z w ])) (12)

[0052] Step (4.2) uses label smoothing during the training process, which is a method of adding noise to the output Y to constrain the model and reduce model overfitting;

[0053] The label smoothing formula in step (4.2) is as follows:

[0054]

[0055]

[0056]

[0057] Where: P i is the probability that the current sample belongs to category i;

[0058] z i Refers to the logit of the corresponding category i of the current sample;

[0059] y i The true label of the sample is in the form of a one-hot vector, assuming that the label has a probability of ∈ that it is incorrectly labeled;

[0060] Step (4.3) Build the EfficientNet model and construct the MBCABlock to improve the network modeling capability;

[0061] Step (4.4) feeds the processed training set data into the network for training.

[0062] The specific process of step (4.4) is as follows:

[0063] Step (4.4.1) Set the training parameter batch_size. The parameter batch_size is the number of images fed into the network at one training iteration. Here, batch_size is set to 4.

[0064] Step (4.4.2) Set the training parameter learning_rate. The parameter learning_rate is a hyperparameter that determines the stride, time, and final convergence of the model during model training. Here, learning_rate is set to 0.01.

[0065] Step (4.4.3) Set the training parameter epochs. The parameter epochs is the total number of times the model is trained to facilitate all images. Here, epochs is set to 200.

[0066] Step (4.4.4) training parameter setting classes_num, the parameter classes_num is the number of model classifications, here the fatty liver B-ultrasound images are classified into 4 categories, and the parameter classes_num is set to 4;

[0067] Step (4.4.5) Set the training parameter momentum. The parameter momentum is a hyperparameter that avoids some uncertainties and accelerates convergence during optimization during training. Here, momentum is set to 0.9.

[0068] Step (4.4.6) sets the training parameter weight_decay. The parameter weight_decay is a penalty term used during training to optimize the model to avoid overfitting. Here, weight_decay is set to 0.0001.

[0069] As a further limitation of the present technical solution, the specific operation of step (5) is as follows:

[0070] Step (5.1) saves the trained model parameters and reads the model parameters during prediction

[0071] Step (5.2) predicts the fatty liver B-ultrasound images in the test set and evaluates the performance of the model under various evaluation indicators.

[0072] Compared with the prior art, the advantages and positive effects of the present invention are:

[0073] 1. Through the implementation and application of this method, the medical clinical diagnosis of fatty liver disease can be significantly improved, and compared with other existing methods, this method has great improvements in model training, inference speed, parameter size, and classification accuracy.

[0074] 2. By comparing with other existing methods, the AdaptiveCoordinateAttention efficient convolutional neural network based on label smoothing has a classification accuracy of up to 94.96% and an F1-Score of 0.955, which is significantly improved compared with network structures such as ResNet, InceptionV4, ResNeXt, Xception, and DenseNet, and the evaluation performance is improved by 11.7%.

[0075] 3. In this method, a series of preprocessing operations such as adaptive histogram equalization, grayscale linear transformation, ROI selection, average grayscale ratio comparison, etc. are performed on the fatty liver B-ultrasound image to achieve data enhancement and feature analysis of the fatty liver B-ultrasound image, which provides a reference for the subsequent design of the network structure. In addition, an AdaptiveCoordinateAttention efficient convolutional neural network based on label smoothing is proposed to reduce overfitting through label smoothing, and can capture long-range dependencies and retain accurate location information along two spaces to improve the modeling ability of the model. This makes the performance of this method superior, and the recognition ability has surpassed manual recognition.

[0076] 4. This method is a lightweight network with fast training and inference, small model parameters, and based on adaptive coordinated attention, it can adaptively encode the horizontal and vertical position information as channel attention, while retaining the position information of other spatial directions to enhance the representation of the object of attention. Based on this method, the recognition ability of fatty liver B-ultrasound images has exceeded manual recognition and is superior to other methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a schematic diagram of the Adaptive Coordinate Attention structure of the present invention.

[0078] Figure 2 It is a schematic diagram of the MBCABlock structure of the present invention.

[0079] Figure 3 It is a flow chart for implementing the present invention. DETAILED DESCRIPTION

[0080] A specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific implementation.

[0081] like Figure 1-Figure 3 As shown, the present invention comprises the following steps:

[0082] Step (1) preprocessing the obtained fatty liver B-ultrasound image to obtain a data enhanced B-ultrasound image;

[0083] In step (1), a series of preprocessing operations are performed on the fatty liver B-ultrasound image, using adaptive grayscale histogram equalization, grayscale linear transformation, and Gaussian smoothing noise reduction, as follows:

[0084] Step (1.1) adjusts the image size from an irregular size to 300*300, and at the same time improves the image resolution;

[0085] Step (1.2) grayscale processing is performed on the image, grayscale distribution is counted, and an algorithm is used to adaptively equalize grayscale histograms, so as to improve near-field and far-field contrast in image display;

[0086] r and s represent the grayscale of the original image and the grayscale of the image after histogram equalization, respectively.

[0087] The formula for adaptive grayscale histogram equalization is as follows:

[0088] s=T(r) r=T -1 (s) (1)

[0089]

[0090]

[0091]

[0092]

[0093] Step (1.3) performs a grayscale linear transformation on the image after histogram equalization. Since the original image has noise blur and various impurity noises, this method is used to highlight the required features while reducing irrelevant features.

[0094] The grayscale linear transformation formula in step (1.3) is as follows:

[0095]

[0096] Where: f(x, y) is the gray value of the input image;

[0097] g(x, y) is the gray value of the output image;

[0098] [a, b] is the grayscale value range of the input image;

[0099] [c, d] is the grayscale value range of the output image;

[0100] By adjusting the intercept of the function on the Y axis, the grayscale values ​​of all pixels are moved up or down, so as to make the entire image darker or brighter.

[0101] Step (1.4) performs Gaussian smoothing noise reduction on the processed image. Gaussian filtering is a linear smoothing filter that has a good suppressive effect on noise that obeys a normal distribution.

[0102] Step (2) extracting the Region Of Interest of the fatty liver B-ultrasound image, performing far-field and near-field segmentation, and performing far-field and near-field grayscale analysis;

[0103] The step (2) specifically includes the following operations:

[0104] Step (2.1) performs near-field and far-field segmentation on the processed image, segmenting the fatty liver B-ultrasound image into near field, mid-field and far field, and extracting two ROIs from each B-ultrasound image, one in the near field and the other in the far field;

[0105] Step (2.2) calculates the average grayscale of the extracted ROI, which can reflect many details of the image;

[0106] The average grayscale calculation formula in step (2.2) is as follows:

[0107]

[0108]

[0109] Where: n k is the grayscale z k The number of times it appears in the image;

[0110] MN is the total number of pixels;

[0111] Step (2.3) performs MIR feature calculation on the calculated ROI average grayscale of different field areas and calculates the near-field and far-field average grayscale ratio. According to the characteristics of near-field echo enhancement and far-field echo attenuation of B-ultrasound images of fatty liver, the selection of MIR features can well reflect the severity of fatty liver.

[0112] The MIR feature calculation formula in step (2.3) is as follows:

[0113]

[0114]

[0115] Where: H and W are the height and width of ROI respectively;

[0116] In this paper, it is a fixed value of 100, x and y represent the coordinates of the pixel point, and I(x, y) represents the grayscale value of the pixel point in the ROI.

[0117] Step (3) performing data segmentation on the preprocessed B-ultrasound image;

[0118] The specific operation of step (3) is as follows:

[0119] 70% of the data set of each type of image is randomly selected as the training set, 20% of the images are used as the validation set, and the remaining 10% are used as the test set. The test set is used to evaluate the model's ability to discriminate new samples, and the test error is used as an approximation of the generalization error.

[0120] Step (4) Design the Adaptive Coordinate Attention structure, use label smoothing, rebuild the EfficientNet model, construct MBCABlock to improve the network modeling capability, feed the preprocessed training set images into the network for training, and adjust the network structure and hyperparameters through the validation set;

[0121] The specific operation of step (4) is as follows:

[0122] Step (4.1) Design the Adaptive Coordinate Attention structure to adaptively encode the horizontal and vertical position information of the image into the channel attention. While capturing long-range dependencies along one spatial direction, it can also retain precise position information along the other spatial direction. Adaptive learning is performed on the information in the two spatial directions to obtain different degrees of weight information.

[0123] Step (4.1) Adaptive Coordinate Attention specific implementation formula is as follows:

[0124]

[0125] f=δ(F1([z h , z w )])) (12)

[0126] Step (4.2) uses label smoothing during the training process, which is a method of adding noise to the output Y to constrain the model and reduce model overfitting;

[0127] The label smoothing formula in step (4.2) is as follows:

[0128]

[0129]

[0130]

[0131] Where: P i is the probability that the current sample belongs to category i;

[0132] z i Refers to the logit of the corresponding category i of the current sample;

[0133] y i The true label of the sample is in the form of a one-hot vector, assuming that the label has a probability of ∈ that it is incorrectly labeled;

[0134] Step (4.3) Build the EfficientNet model and construct the MBCABlock to improve the network modeling capability;

[0135] Step (4.3.1) builds the EfficientNet model, which consists of 9 stages, 1 stem layer, 7 MBCABlock layers, and 1 head layer. The 9 stages are stacked in the above order.

[0136] Step (4.3.2) constructs the stem layer, which includes a stack of convolutional layers, BN layers, and activation layers, mainly to downsample the input image resolution;

[0137] Step (4.3.3) constructs the MBCABlock layer, which is upgraded by 11 Conv, and the DepwiseConv is deeply separable for 2D convolution. The channel attention is superimposed through the SE module, and then a 11 Conv is used for dimensionality reduction and compression to form a bottleneck structure. Then, a Dropout layer is used to prevent overfitting, and finally a residual structure connection is performed;

[0138] Step (4.3.4) constructs the head layer, which is dimensionalized by 11 Conv and BN layers, and constructs a fully connected network output classification category through the AdptiveAvgpool layer, Dropout layer, and Linear layer.

[0139] Step (4.4) feeds the processed training set data into the network for training.

[0140] The specific process of step (4.4) is as follows:

[0141] Step (4.4.1) Set the training parameter batch_size. The parameter batch_size is the number of images fed into the network at one training iteration. Here, batch_size is set to 4.

[0142] Step (4.4.2) Set the training parameter learning_rate. The parameter learning_rate is a hyperparameter that determines the stride, time, and final convergence of the model during model training. Here, learning_rate is set to 0.01.

[0143] Step (4.4.3) Set the training parameter epochs. The parameter epochs is the total number of times the model is trained to facilitate all images. Here, epochs is set to 200.

[0144] Step (4.4.4) training parameter setting classes_num, the parameter classes_num is the number of model classifications, here the fatty liver B-ultrasound images are classified into 4 categories, and the parameter classes_num is set to 4;

[0145] Step (4.4.5) Set the training parameter momentum. The parameter momentum is a hyperparameter that avoids some uncertainties and accelerates convergence during optimization during training. Here, momentum is set to 0.9.

[0146] Step (4.4.6) sets the training parameter weight_decay. The parameter weight_decay is a penalty term used during training to optimize the model to avoid overfitting. Here, weight_decay is set to 0.0001.

[0147] Step (5) uses the trained model to predict images that the model has never seen, namely the test set data, to test the generalization ability of the method.

[0148] The specific operation of step (5) is as follows:

[0149] Step (5.1) saves the trained model parameters and reads the model parameters during prediction

[0150] Step (5.2) predicts the fatty liver B-ultrasound images in the test set and evaluates the performance of the model under various evaluation indicators.

[0151] The above disclosure is only a specific embodiment of the present invention, but the present invention is not limited thereto, and any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for classifying fatty liver ultrasound images using an adaptive coordinated attention network, characterized in that: The following steps are involved: Step (1) preprocessing the obtained fatty liver B-ultrasound image to obtain a data enhanced B-ultrasound image, wherein the preprocessing includes adaptive grayscale histogram equalization, grayscale linear transformation, and Gaussian smoothing noise reduction; Step (2) extracting the Region Of Interest (ROI) of the data-enhanced B-ultrasound image, performing far-field and near-field segmentation, and performing far-field and near-field grayscale analysis; The step (2) specifically includes the following operations: Step (2.1) performs near-field and far-field segmentation on the data-enhanced B-ultrasound image, segmenting the data-enhanced B-ultrasound image into near-field, mid-field and far-field, and extracting two ROIs from each data-enhanced B-ultrasound image, one in the near-field and the other in the far-field; Step (2.2) calculates the average grayscale for the extracted ROI; The average grayscale calculation formula in step (2.2) is as follows: （7） （8） in: is the number of times the grayscale appears in the image; is the total number of pixels; Step (2.3) performs MIR feature calculation on the calculated ROI average grayscale of different field areas and calculates the near-field and far-field average grayscale ratio. According to the characteristics of near-field echo enhancement and far-field echo attenuation of B-ultrasound images, the severity of fatty liver is reflected by selecting MIR features; The MIR feature calculation formula in step (2.3) is as follows: （9） （10） Where: H and W are the height and width of ROI, which are fixed values ​​of 100; x and y are the coordinates of the pixel points, Represents the gray value of the pixel in ROI; Step (3) performing data segmentation on the data-enhanced B-ultrasound images to obtain training set images, validation set images, and test set images; Step (4) Design an Adaptive Coordinate Attention network, use label smoothing, rebuild the EfficientNet network, construct MBCABlock to improve the network modeling capability, input the training set images into the Adaptive Coordinate Attention network for training, and adjust the Adaptive Coordinate Attention network structure and hyperparameters through the validation set images; The specific operation of step (4) is as follows: Step (4.1) Design an Adaptive Coordinate Attention network to adaptively encode the horizontal and vertical position information of the image into the channel attention, while capturing long-range dependencies along one spatial direction, retaining precise position information along the other spatial direction, and adaptively learning the information in the two spatial directions to obtain different degrees of weight information; The specific implementation formula of the Adaptive Coordinate Attention network in step (4.1) is as follows: （11） （12） Step (4.2) uses label smoothing during training and adds noise to the output Y to constrain the adaptive coordinated attention network and reduce the overfitting of the adaptive coordinated attention network; The label smoothing formula in step (4.2) is as follows: （13） （14） （15） in: The current sample belongs to the category probability; Refers to the corresponding category of the current sample of ; is the true label of the sample in the form of a one-hot vector. Assume that the true label of the sample is The probability of labeling errors; Step (4.3) Build the EfficientNet network and construct the MBCABlock to improve the network modeling capabilities; Step (4.4) inputs the training set images into the adaptive coordinated attention network for training; Step (5) uses the trained adaptive coordinated attention network to predict images that have never been seen before, namely, the test set images, to test the generalization ability of the adaptive coordinated attention network for the classification of fatty liver B-ultrasound images.

2. The method for classifying fatty liver B-ultrasound images using an adaptive coordinated attention network according to claim 1, characterized in that: In step (1), the preprocessing of the fatty liver B-ultrasound image is specifically as follows: Step (1.1) adjusting the size of the fatty liver ultrasound image from an irregular size to 300*300 to improve the image resolution; Step (1.2) grayscale processing is performed on the resized fatty liver B-ultrasound image, grayscale distribution is statistically analyzed, and adaptive grayscale histogram equalization is used to improve near-field and far-field contrast in image display; r and s Respectively represent the grayscale of the original image and the grayscale of the image after histogram equalization, The formula for adaptive grayscale histogram equalization is as follows: （1） （2） （3） （4） （5） Step (1.3) performs a grayscale linear transformation on the image after histogram equalization. Since the original image contains noise blur and various impurity noises, the grayscale linear transformation is used to highlight the required features while reducing irrelevant features. The grayscale linear transformation formula in step (1.3) is as follows: （6） in: is the gray value of the input image; is the gray value of the output image; is the gray value range of the input image; is the gray value range of the output image; By adjusting The intercept of the function on the Y axis moves the grayscale value of all pixels up or down, making the entire image darker or brighter. Step (1.4) performs Gaussian smoothing and noise reduction on the processed image.

3. The method for classifying fatty liver ultrasound images using an adaptive coordinated attention network according to claim 1, characterized in that: The specific operation of step (3) is as follows: 70% of the data set of each type of data-augmented B-ultrasound images are randomly selected as training set images, 20% of the data-augmented B-ultrasound images are used as validation set images, and the remaining 10% are used as test set images. The test set images are used to evaluate the discrimination ability of the adaptive coordinated attention network for new samples, and the test error is used as an approximation of the generalization error.

4. The method for classifying fatty liver B-ultrasound images using an adaptive coordinated attention network according to claim 1, characterized in that: The specific process of step (4.4) is as follows: Step (4.4.1) Set the training parameter batch_size. The parameter batch_size is the number of images input into the adaptive coordinated attention network at one time during training iteration. Here, batch_size is set to 4. Step (4.4.2) Set the training parameter learning_rate. The parameter learning_rate is a hyperparameter that determines the optimization stride, time, and final convergence of the adaptive coordinated attention network during training. Here, learning_rate is set to 0.

01. Step (4.4.3) Set the training parameter epochs. The parameter epochs is the total number of times all images are traversed during the training of the adaptive coordinated attention network. Here, epochs is set to 200. Step (4.4.4) training parameter setting classes_num, the parameter classes_num is the number of classifications of the adaptive coordinated attention network. Here, the fatty liver B-ultrasound images are classified into 4 categories, and the parameter classes_num is set to 4; Step (4.4.5) Set the training parameter momentum. The parameter momentum is a hyperparameter that avoids some uncertainties and accelerates convergence during optimization during training. Here, momentum is set to 0.

9. Step (4.4.6) sets the training parameter weight_decay. The parameter weight_decay is a penalty term used to optimize the adaptive coordinated attention network to avoid overfitting during training. Here, weight_decay is set to 0.0001.

5. The method for classifying fatty liver B-ultrasound images using an adaptive coordinated attention network according to claim 1, characterized in that: The specific operation of step (5) is as follows: Step (5.1) saves the trained adaptive coordinated attention network parameters and reads the adaptive coordinated attention network parameters during prediction; Step (5.2) predicts the test set images and evaluates the performance of the adaptive coordinated attention network under various evaluation metrics.

Images