Image simplification method and system for underwater target detection

Abstract

The invention provides an underwater target detection-oriented image simplification method and system, and the method comprises the steps: S1, defining a simplified image compression standard and a feature compression standard, and setting a utility evaluation score threshold; s2, performing feature extraction on the input image, and simplifying the input image and features according to an image and feature compression standard; s3, carrying out information source coding on the simplified image and the features, and then carrying out image and feature utility evaluation; s4, if the utility score is higher than the utility evaluation score threshold, directly performing channel coding on the information source coding image and the features; if the utility evaluation score is lower than the threshold value, the compression is cleared, the compression degree is reduced step by step, the input image and the features are simplified again, and the step S3 is executed again; and S5, decoding at a receiving end. According to the method, effective utility evaluation can be carried out on the simplified image, lightweight processing is carried out on the algorithm, and effective guidance and solutions are provided for the applicability of an image simplification strategy, strategy adjustment and algorithm lightweight.

Description

Technical Field

[0001] The present invention relates to the field of underwater image processing, and in particular to an image reduction method and system for underwater target detection. Background Art

[0002] Underwater images have wide applications and important values in the marine field. In communication in the atmospheric environment, electromagnetic wave communication and optical communication are the main data transmission methods. However, in the harsh underwater environment, the attenuation of electromagnetic waves is relatively large, and the underwater acoustic channel becomes a more suitable choice. However, the extremely narrow bandwidth of the underwater acoustic channel limits the transmission of multimedia data such as images. Before source coding, compression coding is usually used to compress data to improve the transmission efficiency. In view of the fact that underwater images often have practical application backgrounds and are not just for viewing purposes, data reduction (such as the following feature coding, etc.) can be performed on the transmitted images without affecting the execution of tasks, and the amount of data to be transmitted can be further reduced. In this process, how to select an appropriate reduction strategy to ensure that the data is still available in downstream tasks is crucial. Traditional image quality evaluation methods often focus on the quality of images, which emphasize the objective attributes and characteristics of the images themselves, such as resolution, contrast, color restoration, noise level, etc., and measure the quality of images through various technical indicators. However, in the actual application of underwater images, more emphasis should be placed on whether the images can meet the actual needs of users in specific application scenarios. There is currently no complete solution for how to ensure that the data is still available in downstream tasks during the reduction process.

[0003] The present invention aims to design an image reduction method and system for underwater target detection, with the focus on evaluating the utility of the reduced data, using the utility evaluation results to measure the applicability of the reduction strategy and provide directions for strategy adjustment, and lightweight processing of the algorithm to provide conditions for algorithm transplantation. Summary of the Invention

[0004] The purpose of the present invention is to propose an image reduction method and system for underwater target detection. After reducing the image, the utility score of the image is evaluated, and the data can be maximally reduced while maintaining a high utility score, improving the channel transmission efficiency. And lightweight processing of the utility evaluation algorithm is performed to provide conditions for algorithm transplantation.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] The present invention proposes an image reduction method for underwater target detection, including the following steps:

[0007] Step S1, formulate data reduction criteria, define the image compression criteria and feature compression criteria for reduction, and set the threshold of the utility evaluation score;

[0008] Step S2: Extract features from the input image, and streamline the input image and the input image features according to the defined streamlined image compression standard and feature compression standard;

[0009] Step S3: Perform source coding on the streamlined image and features, and respectively evaluate the image and feature utility of the encoded image and features;

[0010] Step S4: If the image and feature utility scores of the encoded image are higher than the utility evaluation score threshold, directly perform channel coding on the source-coded image and features; if the image and feature utility evaluation scores are lower than the threshold, clear the compression, gradually reduce the compression degree, and re-streamline the input image and the input image features, then return to execute Step S3 until the image and feature utility scores of the encoded image and features are higher than the threshold, and then perform channel coding;

[0011] Step S5: Decode at the receiving end to receive the streamlined image and features.

[0012] Preferably, the evaluation of the image and feature utility of the encoded image and features respectively is implemented through the following steps:

[0013] S31: Construct an underwater fish detection dataset FODD; and perform image compression on the FODD dataset as the image part data of the utility dataset UDCUI for streamlined underwater images;

[0014] S32: Select m target detection models, perform the first training on the m target detection models using the COCO dataset to improve the generalization ability of the models, and perform the second training on the models completed in the first training using the FODD dataset to obtain m fish detection models;

[0015] S33: Input the images in the UDCUI dataset into the m fish detection models one by one to obtain m groups of confidence scores, calculate the data saturation, determine the number J of models at data saturation according to the calculation result of the data saturation, and select the average confidence of the J fish detection models from the m fish detection models as the utility score Score of the i-th image i ;

[0016] S34: Each detection model in the J fish detection models selects a corresponding feature compression model; select a group from the J groups of fish detection models and the corresponding feature compression models, and construct the feature part data of the UDCUI dataset according to the selected fish detection model, feature compression model, and the images in the UDCUI dataset; the feature part data includes the compressed features of the i-th image and the confidence score Class_conf corresponding to the compressed features of the i-th image i ;

[0017] S35. Construct an image utility evaluation module based on the J fish detection models selected according to S33 and complete the training using the UDCUI dataset; use the image utility evaluation module as the teacher network for the image part, use ResNet18 as the student network, and lightweight the image utility evaluation module;

[0018] S36. Construct a feature utility evaluation module based on the fish detection model selected according to S34 and complete the training using the UDCUI dataset; use the feature utility evaluation module as the teacher network for the feature part, use ResNet18 as the student network, and lightweight the feature utility evaluation module;

[0019] S37. Input the encoded image and features into the lightweight image utility evaluation module and the lightweight feature utility evaluation module to perform image and feature utility evaluations.

[0020] Preferably, construct an underwater fish detection dataset FODD, which specifically includes the following steps:

[0021] Extract images from consecutive frames of underwater real-time monitoring videos and perform data annotation on the images to obtain information on the types and position coordinates of fish, thus constituting the underwater fish detection dataset FODD for facilitating the training of the detection model.

[0022] Preferably, the specific content of S33 is as follows:

[0023] Start incrementing from the model number J0 = 2 in sequence, calculate the data saturation of different model numbers until the data is saturated, and determine the model number J when the data is saturated. The maximum value of J0 is m;

[0024] Among them, for the calculation of the data saturation of the model number J0, randomly select J0 groups from m groups of confidence scores to obtain kinds of results, each result contains J0 groups of confidence scores, and then calculate the average confidence of each image in these J0 groups of confidences. Therefore, kinds of average confidence results can be obtained. For these kinds of results, randomly select J0 groups of results again through the method of permutation and combination to perform PLCC calculations to obtain n PLCC calculation results; average the n PLCC results to obtain the final result, which is used as the data saturation result when the number of algorithms is J0;

[0025] Comprehensively consider the task background, accuracy, and computational complexity to select J fish detection models from m fish detection models by trade-off, and use the average confidence of the i-th image of the J fish detection models as the utility score of the i-th image;

[0026]

[0027] wherein represents the confidence result predicted by the j-th model for the i-th image.

[0028] Preferably, one group is selected from the J groups of fish detection models and the corresponding feature compression models, and the feature part data of the UDCUI dataset is constructed according to the selected fish detection model, feature compression model, and the images in the UDCUI dataset; specifically as follows:

[0029] The J fish detection models respectively perform feature extraction on the images in the FODD dataset, and the features extracted by each fish detection model are respectively used for the training of the corresponding feature compression model to complete the training of the J feature compression models;

[0030] The images in the UDCUI dataset are sequentially input into the J feature compression models to obtain the compressed features; the obtained compressed features are respectively input into the corresponding fish detection models to obtain J groups of confidence scores;

[0031] Calculate the PLCC of each group of confidence scores and the average value of the J groups of confidence scores respectively, and select the group of models with the highest correlation as the basis for constructing the final compressed features and formulating labels;

[0032] Take the compressed features corresponding to the selected feature compression model as the feature part of the UDCUI dataset, and take the confidence scores of the corresponding fish detection model as the label Label of the compressed features i :

[0033] Label i = Class_conf i × 100,

[0034] where i represents the feature serial number corresponding to the i-th image, and Class_conf i is the confidence score of the i-th feature in the UDUCI dataset.

[0035] Preferably, an image utility evaluation module is constructed according to the J fish detection models selected in S33 and trained using the UDCUI dataset, specifically including the following steps:

[0036] Step A1: Sequentially input the images in the UDCUI dataset into the J fish detection models, and take the features output by each fish detection model as the migrated utility features represents a feature of a certain scale in, and n represents the number of different scales of the features. This process is as follows:

[0037]

[0038] Among them, I i is the i-th input image, and Detection j represents the j-th fish detection model;

[0039] Step A2: Upsample the low-scale features in the utility features to the intermediate scale, downsample the high-scale features to the intermediate scale, and do not perform additional processing on the intermediate-scale features. Concatenate multiple features of the same scale in the channel dimension to obtain a feature group:

[0040]

[0041] Flatten the feature group P i j in the width and height dimensions to obtain a two-dimensional matrix M i j ;

[0042] Step A3: Use positional encoding to record the position data of each feature vector in the two-dimensional matrix, that is, add a learnable position matrix POS to the M i j matrix, and add a vector f i j at the initial end of the matrix M r , expressed in the following form:

[0043]

[0044] Step A4: Each fish detection model is equipped with a corresponding Transformer Encoder. Input into the Transformer Encoder corresponding to the j-th fish detection model, and the mathematical expression of the Transformer Encoder process is as follows:

[0045]

[0046] In the formula, represents the vector in , represents the output of the Transformer Encoder, LN represents layer normalization, MSA represents multi-head self-attention mechanism, and MLP represents a multi-layer perceptron, including two fully connected layers and a GELU activation function;

[0047] After passing through the fully connected layer FC, the utility score

[0048]

[0049] Step A5: Use the mean absolute error MAE as the loss function L I Perform the training of the Transformer Encoder, and the formula is as follows:

[0050]

[0051] where represents the confidence score of the i-th image output by the j-th model;

[0052] Step A6: Fuse the features of J hidden layers through a multi-layer perceptron and output the utility score U i :

[0053]

[0054] The multi-layer perceptron includes two fully connected layers FC1, FC2 and a LeakyReLU activation function;

[0055] Step A7: Perform model training through the loss function L a The calculation method is as follows:

[0056] L a = |Score i - U i |,

[0057] Score i is the utility score of the i-th image in the UDCUI dataset;

[0058] Step A8: Design a common feature extraction module based on the generative adversarial network GAN. During the training process, define the output of the Transformer Encoder as the input feature where i and j respectively represent the i-th image and the j-th model, the type label of is 1, and the generator generates the feature the type label is 0; the generator consists of two modules FC'1 and FC'2, and each module contains a fully connected layer and a LeakyReLU activation function. First, the feature generates the preliminary common feature

[0059]

[0060] Step A9: The discriminator consists of three modules, FC3, FC4, and FC5. FC3 contains two fully connected layers, the LeakyReLU, and the Sigmoid activation functions; the features After passing through the FC3 module, the output of each neuron of the features is mapped to the range from 0 to 1:

[0061]

[0062] where are the features output by FC3;

[0063] Step A10: Output the class probability distribution and the type probability distribution through the FC4 and FC5 modules. Among them, FC4 is for multi-classification tasks and contains one fully connected layer and the Softmax activation function; FC5 is for binary classification tasks and contains one fully connected layer and the Sigmoid activation function:

[0064]

[0065] Step A11: Define the loss function during the training process as:

[0066]

[0067] where L S and L V respectively represent the log-likelihoods of the correct type label and the correct class label, and thus obtain the final loss function L T formula:

[0068]

[0069] where the training objectives of the generator and the discriminator are to minimize L S and L V and to maximize L S and L V . Through iterative training, the generator extracts the features that the discriminator cannot correctly classify as common features

[0070] Step A12: Use the principal component analysis (PCA) transformation to generate the common features generated by the generator for data alignment and remove the specific data in the common features; splice the J features from to to obtain the input f i of the PCA, and de-mean f i to get The mathematical expression of this process is as follows:

[0071]

[0072] where is the mean vector of f i ;

[0073] Step A13: Calculate the covariance matrix Cov of :

[0074]

[0075] where D is the size of the first dimension of ; is the transpose matrix of ; Perform eigenvalue decomposition on Cov:

[0076] Cov·V = V·Λ,

[0077] where V is the matrix composed of eigenvectors, Λ is the diagonal matrix with eigenvalues on the diagonal, and each column eigenvector in V corresponds to the eigenvalue in Λ; Arrange the eigenvalues from largest to smallest, and select the eigenvectors corresponding to the first k eigenvalues to form the principal component projection matrix P(v1, v2…v k );

[0078] Step A14: Project the de-meaned matrix onto the principal component projection matrix P to obtain the aligned feature Y i :

[0079]

[0080] The aligned feature Y i is used for quality prediction through FC6. FC6 contains two fully connected layers and an activation function LeakyReLU to obtain the utility score U i ':

[0081] U' i = FC6(Y i ),

[0082] Step A15: Define the loss function L f during training as follows:

[0083] L f = |Score i - U i '|,

[0084] Score i is the utility score of the i-th image in the UDCUI dataset.

[0085] Preferably, the image utility evaluation module is used as the teacher network of the image part, and ResNet18 is used as the student network to lightweight the image utility evaluation module, which is specifically as follows:

[0086] Step A16: Pass the input image I i through a convolution Conv1 for shallow feature extraction, and then obtain the shallow feature F through max pooling Maxpool e , and the shallow feature F e successively passes through M Block layers for enhanced feature extraction to obtain the deep feature F s :

[0087] F e = Maxpool(Conv1(I i )),

[0088] F s = BlockM(...(Block2(Block1(F e )))),

[0089] Step A17: After being processed by average pooling Avgpool and the fully connected layer FC″1, obtain the hidden layer feature l i , and the hidden layer feature l i will be mapped into the image utility score S after passing through the fully connected layer FC″2 i :

[0090] l i = FC″1(Avgpool(F s )),

[0091] S i = FC″2(l i ),

[0092] Step A18: During training, the lightweight image utility evaluation module constructs a loss function I based on MAE total , which consists of two parts, including the feature loss I loss and the utility score loss I s , and the calculation process is as follows:

[0093] I total = I loss + I s ,

[0094] I loss = |l - l i |,

[0095] I s = |Score i - S i |,

[0096] Among them, l is the feature of the hidden layer of the teacher network, that is, the aligned feature Y i .

[0097] Preferably, a feature utility evaluation module is constructed according to the fish detection model selected in S34 and trained using the UDCUI dataset, which specifically includes the following steps:

[0098] Step B1: Modify the selected fish detection model so that the input of the fish detection model can match the dimension of the features in the UDCUI dataset;

[0099] Step B2: Input the features in the UDCUI dataset into the fish detection model obtained in Step B1 in sequence, and use the output features as the migrated utility features F i = [f i,1 ; f i,2 ; …; f i,n , f i,n represents a feature of a certain scale in F i , n represents the number of different scales of the features, and this process is as follows:

[0100] F i = Detection(f i ),

[0101] where f i is the feature corresponding to the i-th image in the input UDCUI dataset, and Detection represents the fish detection model obtained in Step B1;

[0102] Step B3: Upsample the low-scale features in the utility feature F i to the intermediate scale, downsample the high-scale features to the intermediate scale, and do not perform additional processing on the intermediate-scale features. Concatenate multiple features of the same scale in the channel dimension to obtain a feature group:

[0103] P i = [f′ i,1 ; f′ i,2 …; f′ i,n ,

[0104] Flatten the feature group P i in the width and height dimensions to obtain a two-dimensional matrix M i ;

[0105] Step B4: Use the position encoding method to record the position data of each feature vector in the two-dimensional matrix, that is, add a learnable position matrix POS to the M i matrix, and in the matrix M iAdd the vector f to the initial end r , expressed in the following form:

[0106] M′ i =[f i,r ; f′ i,1 ; f′ i,2 …; f′ i,n +POS,

[0107] Step B5. The fish detection model obtained in Step B1 is provided with a corresponding Transformer Encoder. Input M″ i into the Transformer Encoder. The mathematical expression of M″ i through the Transformer Encoder process is as follows:

[0108] f″′ i,l =MSA(LN(f″′ i,l-1 ))+f″′ i,l-1 , l = 1…n,

[0109] f′ i,l =MLP(LN(f″ i,l ))+f″ i,l ,

[0110] In the formula, f″′ i,l-1 represents the vector in M′ i , f′ i,l represents the output of the TransformerEncoder, and f′ i,l obtains the utility score S i through the fully connected layer FC:

[0111] S i =FC(f′ i,l ),

[0112] Step B6. Use the mean absolute error as the loss function L′ I to train the TransformerEncoder. The formula is as follows:

[0113] L′ I =|Class_conf i -S i |,

[0114] where Class_conf i is the confidence score of the i-th feature in the UDUCI dataset;

[0115] Step B7. The output features f′ i,l, through a multi-layer perceptron, and output the final utility score U″ i :

[0116] U″ i = FC2(LeakyReLU(FC1(f′ i,l )))

[0117] The multi-layer perceptron includes two fully connected layers FC1, FC2 and a LeakyReLU activation function. The model is trained through the loss function L′ a and the calculation method is as follows:

[0118] L′ a = |Label i - U″ i |,

[0119] Label i is the label of the i-th feature in the UDCUI dataset.

[0120] Preferably, the feature utility evaluation module is used as the teacher network of the feature part, and ResNet18 is used as the student network to lightweight the feature utility evaluation module, which specifically includes the following steps:

[0121] Step B8: Input the feature f i , perform shallow feature extraction through the convolutional layer Conv1 to obtain the shallow feature F′ e , and then sequentially perform enhanced feature extraction through M Block layers to obtain the deep feature F′ s :

[0122] F′ e = Conv1(f i ),

[0123] F′ s = BlockM(...(Block2(BlockL(F′ e )))),

[0124] Step B9: After average pooling and processing by the fully connected layer FC″1 layer, obtain the hidden layer feature l f , and the hidden layer feature will be mapped into the feature utility score S f after passing through the fully connected layer FC″2;

[0125] l f = FC″1(Avgpool(F′ s )),

[0126] S f = FC″2(l f ),

[0127] Step B10. During training, the lightweight feature utility evaluation module constructs a loss function F based on the mean absolute error total , which consists of two parts, including the feature loss F loss and the utility score loss F s . The calculation process is as follows:

[0128] F total = F loss + F s ,

[0129] F loss = |l T - l f |,

[0130] F s = |Label i - S f |,

[0131] where l T is the output f′ of the multi-layer perceptron in the feature utility evaluation module i,l ; Label i is the label of the i-th feature in the UDCUI dataset.

[0132] The present invention also proposes an image reduction system for underwater target detection, including a processor, a memory, and a computer program stored on the memory. When the processor executes the computer program, it specifically executes the steps in any of the above-mentioned image reduction methods for underwater target detection.

[0133] Compared with the prior art, the present invention has the following beneficial effects:

[0134] The present invention can effectively evaluate the utility of the reduced image, and lightweight the algorithm, providing effective guidance and solutions for the applicability of the image reduction strategy, strategy adjustment, and algorithm lightweighting. BRIEF DESCRIPTION OF THE DRAWINGS

[0135] Figure 1 is a block diagram of the reduced underwater image utility evaluation system for target detection designed by the present invention;

[0136] Figure 2 is a structural diagram of the utility feature processing of the present invention;

[0137] Figure 3 is a structural diagram of the TransformerEncoder of the present invention;

[0138] Figure 4 is the image utility evaluation algorithm framework of the present invention;

[0139] Figure 5 This is the structure diagram of the common feature extraction module of the present invention;

[0140] Figure 6 This is the structure diagram of the algorithm lightweighting of the present invention;

[0141] Figure 7 This is the framework of the feature utility evaluation algorithm of the present invention. Specific implementation manners

[0142] Next, in combination with the attached Figure 1-7 , the technical solution of the present invention will be specifically described.

[0143] The present invention proposes an image reduction method for underwater target detection, including the following steps:

[0144] Step S1, formulate data reduction criteria, define the image compression criteria and feature compression criteria for reduction, and set the utility evaluation score threshold; in this embodiment, the image compression criteria are defined as: on the basis of 4-fold downsampling, use the JPEG standard with Q = 75 for compression.

[0145] Step S2, perform feature extraction on the input image, and reduce the input image and the input image features according to the defined image compression criteria and feature compression criteria for reduction;

[0146] Step S3, perform source coding on the reduced image and features, and respectively perform image and feature utility evaluations on the encoded image and features;

[0147] Step S4, if the image and feature utility scores of the encoded image are higher than the utility evaluation score threshold, directly perform channel coding on the source-coded image and features; if the image and feature utility evaluation scores are lower than the threshold, clear the compression, gradually reduce the compression degree and re-reduce the input image and the input image features, including gradually reducing the downsampling ratio and gradually increasing the value of Q, and return to execute step S3 until the image and feature utility scores of the encoded image and features are higher than the threshold, and then perform channel coding;

[0148] Step S5, perform decoding at the receiving end to receive the reduced image.

[0149] In this embodiment, the above-mentioned performing image and feature utility evaluations on the encoded image and features respectively is implemented through the following steps:

[0150] S31, construct an underwater fish detection dataset FODD; and perform image compression on the FODD dataset as the image part data of the utility dataset UDCUI (Utility Datasets for Compacted Underwater Image) for the reduced underwater image;

[0151] S32. Select m target detection models. In this embodiment, 11 target detection algorithms are selected for data saturation calculation, including: CenterNet, EfficientDet, YOLOv5, YOLOv6, YOLOX, YOLOv7, YOLOv8, Faster-RCNN, DETR, Swin Transformer, and SSD. Use the COCO dataset to perform the first training on the m target detection models to improve the generalization ability of the models. Use the FODD dataset to perform the second training on the models completed in the first training to obtain m fish detection models.

[0152] S33. Input the images in the UDCUI dataset into the m fish detection models one by one to obtain m sets of confidence scores. In this embodiment, each set contains the confidence scores corresponding to 24,678 images. Calculate the data saturation, and determine the number of models J when the data is saturated according to the calculation result of the data saturation. In this embodiment, the number of models J when the data is saturated is J = 5. Select the average confidence of the J fish detection models from the m fish detection models as the utility score Score of the i-th image. i ;

[0153] S34. For each detection model in the J fish detection models, select a corresponding feature compression model. Select a group from the J groups of fish detection models and the corresponding feature compression models, and construct the feature part data of the UDCUI dataset according to the selected fish detection model, feature compression model, and the images in the UDCUI dataset. The feature part data includes the compressed features of the i-th image and the confidence score Class_conf corresponding to the compressed features of the i-th image. i ;

[0154] S35. Construct an image utility evaluation module according to the J fish detection models selected in S33 and complete the training using the UDCUI dataset. Use the image utility evaluation module as the teacher network for the image part, use ResNet18 as the student network, and lightweight the image utility evaluation module.

[0155] S36. Construct a feature utility evaluation module according to the fish detection model selected in S34 and complete the training using the UDCUI dataset. Use the feature utility evaluation module as the teacher network for the feature part, use ResNet18 as the student network, and lightweight the feature utility evaluation module.

[0156] S37. Input the encoded image and features into the lightweight image utility evaluation module and the lightweight feature utility evaluation module for image and feature utility evaluation.

[0157] In this embodiment, an underwater fish detection dataset FODD is constructed, which specifically includes the following steps:

[0158] Extract images from consecutive frames of an underwater real-time monitoring video, and perform data annotation on the images to obtain information on the types and position coordinates of fish, thereby constituting an underwater fish detection dataset (Fish Object Detection Database, FODD), which is convenient for the training of the detection model.

[0159] In this embodiment, image compression of the FODD dataset refers to performing JPEG compression after downsampling the image. The compression scheme is shown in Table 1:

[0160] Table 1

[0161]

[0162]

[0163] In this embodiment, the specific content of S33 is as follows:

[0164] Starting from the model number J0 = 2 and increasing sequentially, calculate the data saturation of different model numbers until the data is saturated, and determine the model number J when the data is saturated. The maximum value of J0 is m;

[0165] Among them, the calculation of the data saturation of the model number J0 is to randomly select J0 groups from m groups of confidence scores to obtain kinds of results, each result contains J0 groups of confidence scores, and then calculate the average confidence of each picture in these J0 groups of confidence scores. Therefore, kinds of average confidence results can be obtained. For these kinds of results, randomly select J0 groups of results again through permutation and combination method to perform PLCC calculation, and obtain n PLCC calculation results; average the n PLCC results to obtain the final result, which is used as the data saturation result when the number of algorithms is J0;

[0166] Among them, the calculation of the data saturation of the model number J0 is to randomly select J0 groups from m groups of confidence scores, calculate the average confidence of each picture in the selected J0 groups, calculate the Pearson linear correlation coefficient PLCC (Pearson linear correlation coefficient) between each group of confidence scores and the average confidence of the J0 groups, and average the PLCC results to calculate the data saturation result;

[0167] When the number of extracted algorithms is 2, according to the calculation method of permutation and combination There are a total of 55 extraction results, and each result contains two sets of confidence scores. Then, calculate the average confidence of each image in these two sets of confidence scores, so that 55 sets of average confidence results can be obtained. For these 55 sets of results, randomly extract the results of two algorithms again through the method of permutation and combination and calculate their Pearson linear correlation coefficient (PLCC). A total of 1485 PLCC calculation results can be obtained. After averaging these 1485 PLCC results, the final result is obtained as the data saturation result when the number of algorithms is 2;

[0168] When the number of algorithms is 3, the steps are the same as above. Through the calculation of the first permutation and combination, kinds of extraction results are obtained, and each result contains three sets of confidence scores. After taking the average of every three sets of scores, perform the second permutation and combination calculation Repeat the above method and extrapolate to the results when the number of algorithms is 11; in this embodiment, the number of models J at data saturation is 5;

[0169] Comprehensively considering the task background, accuracy, and computational complexity, select J fish detection models from m fish detection models by trade-off, and use the average confidence of the i-th image of the J fish detection models as the utility score of the i-th image;

[0170]

[0171] where represents the confidence result predicted by the j-th model for the i-th image.

[0172] In this embodiment, select a group from the J groups of fish detection models and corresponding feature compression models, and construct the feature part data of the UDCUI dataset according to the selected fish detection model, feature compression model, and the images in the UDCUI dataset; specifically as follows:

[0173] The J fish detection models respectively extract features from the images in the FODD dataset, and the features extracted by each fish detection model are respectively used for the training of the corresponding feature compression model, and the training of the J feature compression models is completed;

[0174] Input the images in the UDCUI dataset into the J feature compression models in turn to obtain the compressed features; input the obtained compressed features into the corresponding fish detection models respectively to obtain J sets of confidence scores;

[0175] Calculate the PLCC of each set of confidence scores and the average of the J sets of confidence scores respectively, and select the group of models with the highest correlation as the basis for constructing the final compressed features and formulating labels;

[0176] Use the compressed features corresponding to the selected feature compression model as the feature part of the UDCUI dataset, and use the confidence score of the corresponding fish detection model as the label Label of the compressed features i :

[0177] Label i = Class_conf i × 100,

[0178] where i represents the feature serial number corresponding to the i-th image.

[0179] In this embodiment, constructing an image utility evaluation module using the J fish detection models selected according to S33 and completing training using the UDCUI dataset specifically includes the following steps:

[0180] Step A1: Sequentially input the images in the UDCUI dataset into the J fish detection models, and use the features output by each fish detection model as the migrated utility features represents the feature of a certain scale in, n represents the number of different scales of the features, and this process is as follows:

[0181]

[0182] where I i is the i-th input image, Detection j represents the j-th fish detection model;

[0183] Step A2: Upsample the low-scale features in the utility features to the intermediate scale, downsample the high-scale features to the intermediate scale, do not perform additional processing on the intermediate-scale features, and splice the features of multiple same scales in the channel dimension to obtain a feature group:

[0184]

[0185] Flatten the feature group P i j in the width and height dimensions to obtain a two-dimensional matrix M i j ;

[0186] Step A3: Use the method of positional encoding to record the position data of each feature vector in the two-dimensional matrix, that is, add a learnable position matrix POS to the M i j matrix, and add a zero vector f of the same length as it at the initial end of the matrix M i j ​r , expressed in the following form:

[0187]

[0188] Step A4. Each fish detection model is equipped with a corresponding Transformer Encoder. Input into the Transformer Encoder corresponding to the j-th fish detection model. The mathematical expression through the Transformer Encoder process is as follows:

[0189]

[0190] In the formula, represents the vector in represents the output of the Transformer Encoder, LN represents layer normalization, MSA represents multi-head self-attention mechanism, and MLP represents multi-layer perceptron, including two fully connected layers and a GELU activation function;

[0191] The utility score

[0192]

[0193] is obtained through the fully connected layer FC. I Step A5. Use the mean absolute error MAE as the loss function L

[0194]

[0195] where represents the confidence score of the i-th image output by the j-th model;

[0196] Step A6. A multi-layer perceptron is used to fuse the features of J hidden layers and output the utility score U i :

[0197]

[0198] The multi-layer perceptron includes two fully connected layers FC1, FC2 and a LeakyReLU activation function;

[0199] Step A7. Model training is carried out through the loss function L a and the calculation method is as follows:

[0200] L a= |Score i -U i |,

[0201] Score i is the utility score of the i-th image in the UDCUI dataset;

[0202] Step A8: Design a common feature extraction module based on the Generative Adversarial Network (GAN). During the training process, define the output of the Transformer Encoder as the input feature where i and j represent the i-th image and the j-th model respectively, and the type label of is 1, and the generator generates the feature with the type label of 0; The generator consists of two modules, FC'1 and FC'2, and each module contains a fully connected layer and a LeakyReLU activation function. First, the feature generates the preliminary common feature through the generator

[0203]

[0204] Step A9: The discriminator consists of three modules, FC3, FC4, and FC5. FC3 contains two fully connected layers, LeakyReLU, and Sigmoid activation functions; The feature passes through the FC3 module, and the output of each neuron of the feature is mapped to the range of 0 to 1:

[0205]

[0206] where is the feature output by FC3;

[0207] Step A10: Output the class probability distribution and the type probability distribution through the FC4 and FC5 modules, where FC4 is for multi-classification tasks and contains a fully connected layer and a Softmax activation function; FC5 is for binary classification tasks and contains a fully connected layer and a Sigmoid activation function:

[0208]

[0209] Step A11: Define the loss function during the training process as:

[0210]

[0211] where LS and L V respectively represent the log-likelihoods of the correct type tags and the correct category tags, from which the final loss function L is obtained T Formula:

[0212]

[0213] where the training objectives of the generator and the discriminator are to minimize L S and L V and to maximize L S and L V . Through iterative training, the generator extracts the features that the discriminator cannot correctly classify as common features

[0214] Step A12: Use the principal component analysis PCA (Principal Component Analysis) to transform the common features generated by the generator for data alignment and remove the specific data in the common features; splice the J features from to to obtain the input f of PCA i , and de-mean f i to get The mathematical expression of this process is as follows:

[0215]

[0216]

[0217] where is the mean vector of f i ;

[0218] Step A13: Calculate the covariance matrix Cov of :

[0219]

[0220] where D is the size of the first dimension of , is the transpose matrix of ; perform eigenvalue decomposition on Cov:

[0221] Cov·V = V·Λ,

[0222] where V is the matrix composed of eigenvectors, Λ is the diagonal matrix with eigenvalues on the diagonal, and each column eigenvector in V corresponds to the eigenvalue in Λ; arrange the eigenvalues from largest to smallest, and select the eigenvectors corresponding to the first k eigenvalues to form the principal component projection matrix P(v1, v2…v k )

[0223] Step A14, the matrix after mean removal Project it into the principal component projection matrix P to obtain the aligned feature Y i :

[0224]

[0225] The aligned feature Y i Perform quality prediction through FC6. FC6 contains two fully connected layers and an activation function LeakyReLU to obtain the utility score U' after common feature extraction i :

[0226] U' i = FC6(Y i ),<![CDATA[ ]]><![CDATA[

[0227] ]]>Step A15, define the loss function L during training f as follows:

[0228] L f = |Score i - U' i |,

[0229] Score i is the utility score of the i-th image in the UDCUI dataset

[0230] In this embodiment, the image utility evaluation module is used as the teacher network of the image part, and ResNet18 is used as the student network to lightweight the image utility evaluation module, specifically as follows:

[0231] Step A16, the input image I i undergoes shallow feature extraction through a convolution Conv1. In this embodiment, a convolution Conv1 with 64 channels and a convolution kernel size of 7×7 is used, and then the shallow feature F is obtained through max pooling Maxpool (MaxPooling) e , the shallow feature F e successively passes through M Block layers for enhanced feature extraction. In this embodiment, 4 Block layers are used to obtain the deep feature F s :

[0232] F e = Maxpool(Conv1(I i )),

[0233] F s = BlockM(...(Block2(Block1(F e )))),

[0234] Step A17: After being processed by the average pooling Avgpool and the fully connected layer FC″1, the hidden layer feature l is obtained i , the hidden layer feature l i will be mapped into the image utility score S after passing through the fully connected layer FC′2′ i :

[0235] l i = FC″1(Avgpool(F s )),

[0236] S i = FC″2(l i ),

[0237] Step A18: During training, the lightweight image utility evaluation module constructs a loss function I based on MAE total , which consists of two parts, including the feature loss I loss and the utility score loss I s , and the calculation process is as follows:

[0238] I total = I loss + I s ,

[0239] I loss = |l - l i |,

[0240] I s = |Score i - S i |,

[0241] where l is the hidden layer feature of the teacher network, that is, the aligned feature Y i .

[0242] In this embodiment, constructing a feature utility evaluation module according to the fish detection model selected in S34 and completing training using the UDCUI dataset specifically includes the following steps:

[0243] Step B1: Modify the selected fish detection model so that the input of the fish detection model can match the dimension of the features in the UDCUI dataset;

[0244] Step B2: Input the features in the UDCUI dataset into the fish detection model obtained in Step B1 in sequence, and use the output features as the migrated utility features F i = [f i,1 ; f i,2 ; …; f i,n , f i,n represents F iFeatures of a certain scale in, n represents the number of different scales of the features, and this process is as follows:

[0245] F i = Detection(f i ),

[0246] where f i is the feature corresponding to the i-th image in the input UDCUI dataset, and Detection represents the fish detection model obtained in step B1;

[0247] Step B3: Upsample the low-scale features in the utility feature F i to the intermediate scale, downsample the high-scale features to the intermediate scale, and do not perform additional processing on the intermediate-scale features. Concatenate multiple features of the same scale in the channel dimension to obtain a feature group:

[0248] P i = [f′ i,1 ; f′ i,2 …; f′ i,n ,

[0249] Flatten the feature group P i in the width and height dimensions to obtain a two-dimensional matrix M i ;

[0250] Step B4: Use positional encoding to record the position data of each feature vector in the two-dimensional matrix, that is, add a learnable position matrix POS to the M i matrix, and add a zero vector f i of the same length as it at the initial end of the matrix M r , expressed in the following form:

[0251] M′ i = [f i,r ; f′ i,1 ; f′ i,2 …; f′ i,n +POS,

[0252] Step B5: The fish detection model obtained in step B1 is set with a corresponding Transformer Encoder. Input M′ i into the Transformer Encoder. The mathematical expression of M′ i passing through the Transformer Encoder is as follows:

[0253] f″ i,l = MSA(LN(f″′ i,l-1 ))+f″′ i,l-1 , l = 1…n,

[0254] f′ i,l = MLP(LN(f″ i,l )) + f″ i,l ,

[0255] where f″′ i,l-1 represents the vector in M′ i , f′ i,l represents the output of the Transformer Encoder, and f′ i,l passes through the fully connected layer FC to obtain the utility score S i :

[0256] S i = FC(f′ i,l ),

[0257] Step B6: Use the mean absolute error as the loss function L′ I to train the Transformer Encoder. The formula is as follows:

[0258] L′ I = |Class_conf i - S i |,

[0259] where Class_conf i is the confidence score of the i-th feature in the UDUCI dataset;

[0260] Step B7: Pass the hidden layer output feature f′ i,l through a multi-layer perceptron and output the final utility score U″ i :

[0261] U″ i = FC2(LeakyReLU(FC1(f″ i,l ))),

[0262] The multi-layer perceptron includes two fully connected layers FC1, FC2 and a LeakyReLU activation function. Train the model using the loss function L′ a The calculation method is as follows:

[0263] L′ a = |Label i - U″ i |,

[0264] Label i is the label of the i-th feature in the UDCUI dataset.

[0265] In this embodiment, the feature utility evaluation module is used as the teacher network of the feature part, and ResNet18 is used as the student network to lightweight the feature utility evaluation module. The specific steps are as follows:

[0266] Step B8: Pass the input feature f i , and perform shallow feature extraction through the convolutional layer Conv1. In this embodiment, a convolutional layer Conv1 with 128 channels and a convolutional kernel of 64×64 is used to obtain the shallow feature F′ e . Subsequently, strengthen the feature extraction through M Block layers in sequence. In this embodiment, 4 Block layers are used to obtain the deep feature F′ s :

[0267] F′ e =Conv1(f i ),

[0268] F′ s =BlockM(...(Block2(Block1(F′ e )))),

[0269] Step B9: After being processed by average pooling and the fully connected layer FC″1, obtain the hidden layer feature l f . The hidden layer feature will be mapped into the feature utility score S after passing through the fully connected layer FC″2 f ;

[0270] l f =FC″1(Avgpool(F′ s )),

[0271] S f =FC″2(l f ),

[0272] Step B10: During training, the lightweight feature utility evaluation module constructs a loss function F total based on the mean absolute error. It consists of two parts, including the feature loss F loss and the utility score loss F s . The calculation process is as follows:

[0273] F total =F loss +F s ,

[0274] F loss =|l T -l f |,

[0275] F s =|Label i -Sf ,

[0276] where l T is the output f′ of the multi-layer perceptron in the feature utility evaluation module i,l ; Label i is the label of the i-th feature in the UDCUI dataset.

[0277] The present invention also provides an image reduction system for underwater target detection, including a processor, a memory, and a computer program stored on the memory. When the processor executes the computer program, it specifically executes the steps in any one of the above-mentioned image reduction methods for underwater target detection.

[0278] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0279] The present application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, so that the instructions executed by the processors of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0280] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implements the functions specified in Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0281] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are executed on the computer or other programmable apparatus to generate a computer-implemented process, thereby the instructions executed on the computer or other programmable apparatus provide steps for realizing the functions specified in one process or a plurality of processes and / or one block or a plurality of blocks. Figure 1 one process or a plurality of processes and / or Figure 1 one block or a plurality of blocks.

[0282] As mentioned above, the above are only the preferred embodiments of the present invention, and are not intended to limit the present invention in any other form. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution content of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. An image reduction method for underwater target detection, characterized in that It includes the following steps: Step S1: Formulate data reduction criteria, define reduced image compression criteria and feature compression criteria, and set the threshold of utility evaluation scores; Step S2: Extract features from the input image, and reduce the input image and input image features according to the defined reduced image compression criteria and feature compression criteria; Step S3: Perform source coding on the reduced image and features, and perform image and feature utility evaluations on the encoded image and features respectively; Step S4: If the image and feature utility scores of the encoded image are higher than the utility evaluation score threshold, directly perform channel coding on the source-coded image and features; if the image and feature utility evaluation scores are lower than the threshold, clear the compression, gradually reduce the compression degree, and re-reduce the input image and input image features, and return to execute Step S3 until the image and feature utility scores of the encoded image and features are higher than the threshold, and then perform channel coding; Step S5: Decode at the receiving end to receive the reduced image and features.

2. An image reduction method for underwater target detection according to claim 1, wherein the image and feature utility evaluations of the encoded image and features are respectively implemented through the following steps: S31: Construct an underwater fish detection dataset FODD; and perform image compression on the FODD dataset as the image part data of the utility dataset UDCUI for reduced underwater images; S32: Select m target detection models, perform the first training on the m target detection models using the COCO dataset, and perform the second training on the models completed in the first training using the FODD dataset to obtain m fish detection models; S33. Input the images in the UDCUI dataset into m fish detection models one by one to obtain m sets of confidence scores, calculate the data saturation, determine the number of models J at data saturation according to the calculation result of the data saturation, and select the average confidence of J fish detection models from the m fish detection models as the utility score Score of the i-th image i ; S34: Each of the J fish detection models selects a corresponding feature compression model; Select a group from the J groups of fish detection models and corresponding feature compression models, and construct the feature part data of the UDCUI dataset according to the selected fish detection model, feature compression model, and the images in the UDCUI dataset; The feature part data includes the compressed features of the i-th image and the confidence score Class_conf corresponding to the compressed features of the i-th image i ; S35: Construct an image utility evaluation module according to the J fish detection models selected in S33 and complete the training using the UDCUI dataset; use the image utility evaluation module as the teacher network for the image part, use ResNet18 as the student network, and lightweight the image utility evaluation module; S36: Construct a feature utility evaluation module according to the fish detection model selected in S34 and complete the training using the UDCUI dataset; use the feature utility evaluation module as the teacher network for the feature part, use ResNet18 as the student network, and lightweight the feature utility evaluation module; S37: Input the encoded image and features into the lightweight image utility evaluation module and lightweight feature utility evaluation module for image and feature utility evaluations.

3. An image reduction method for underwater target detection according to claim 2, characterized in that, Constructing the underwater fish detection dataset FODD specifically includes the following steps: Extract images from consecutive frames of the underwater real-time monitoring video, and perform data annotation on the images to obtain information on the types and position coordinates of fish, thereby constituting the underwater fish detection dataset FODD for facilitating the training of the detection model.

4. An image reduction method for underwater target detection according to claim 2, characterized in that, The specific content of S33 is as follows: Starting from the number of models \(J_0 = 2\) and incrementing sequentially, calculate the data saturation for different numbers of models until the data is saturated, and determine the number of models \(J\) when the data is saturated. The maximum value of \(J_0\) is \(m\). Among them, for the calculation of the data saturation with the number of models being J0, J0 groups are randomly selected from m groups of confidence scores to obtain results, each result containing J0 groups of confidence scores. Then, the average confidence of each image in these J0 groups of confidence is calculated. Therefore, average confidence results can be obtained. For these results, through permutation and combination again, J0 groups of results are randomly selected for PLCC calculation to obtain PLCC calculation results; the PLCC results are averaged to obtain the final result, which is used as the data saturation result when the number of algorithms is J0; Comprehensively considering the task background, accuracy, and computational complexity, select \(J\) fish detection models by trade - off from \(m\) fish detection models, and use the average confidence of the \(i\) - th image of the \(J\) fish detection models as the utility score of the \(i\) - th image. where represents the confidence result predicted by the j-th model for the i-th image.

5. The image reduction method for underwater target detection according to claim 4, characterized in that Select one group from the \(J\) groups of fish detection models and corresponding feature compression models, and construct the feature - part data of the UDCUI dataset according to the selected fish detection model, feature compression model, and images in the UDCUI dataset. Specifically as follows: The \(J\) fish detection models respectively extract features from the images in the FODD dataset, and the features extracted by each fish detection model are respectively used for the training of the corresponding feature compression model to complete the training of the \(J\) feature compression models. Input the images in the UDCUI dataset into the \(J\) feature compression models in sequence to obtain the compressed features; input the obtained compressed features into the corresponding fish detection models respectively to get \(J\) groups of confidence scores. Calculate the PLCC between each group of confidence scores and the average of the \(J\) groups of confidence scores respectively, and select the group of models with the highest correlation as the basis for constructing the final compressed features and formulating labels. Use the compressed features corresponding to the selected feature compression model as the feature part of the UDCUI dataset, and use the confidence score of the corresponding fish detection model as the label Label of the compressed features i : Label i = Class_conf i × 100, where i represents the feature serial number corresponding to the i-th figure, and Class_conf i is the confidence score of the i-th feature in the UDUCI dataset.

6. The image reduction method for underwater target detection according to claim 5, characterized in that Construct an image utility evaluation module according to the \(J\) fish detection models selected in S33 and complete the training using the UDCUI dataset, specifically including the following steps: Step A1: Input the images in the UDCUI dataset into J fish detection models in sequence, and use the features output by each fish detection model as the migrated utility features Represents F i j Features at a certain scale in, n represents the number of different scales of the features. This process is as follows: where I i is the i-th input image, and Detection j represents the j-th fish detection model; Step A2: Upsample the low-scale features in utility feature F i j to the intermediate scale, downsample the high-scale features to the intermediate scale, and do not perform additional processing on the intermediate-scale features. Concatenate the features of multiple identical scales in the channel dimension to obtain a feature group: For the feature group P i j Flatten in the width and height dimensions to obtain a two-dimensional matrix M i j ; Step A3: Record the position data of each eigenvector in the two-dimensional matrix by means of position encoding, that is, add a learnable position matrix POS to the M i j matrix, and add a vector f i j at the initial end of the matrix M r , expressed in the following form: Step A4. Each fish detection model is provided with a corresponding Transformer Encoder, and M′ i j is input into the Transformer Encoder corresponding to the j-th fish detection model, and M′ i j The mathematical expression through the Transformer Encoder process is as follows: In the formula, represents the vector in M i ′ j . represents the output of the Transformer Encoder, LN represents layer normalization, MSA represents multi-head self-attention mechanism, and MLP represents multi-layer perceptron, including two fully connected layers and a GELU activation function; The utility score is obtained through the fully connected layer FC Step A5: Use the Mean Absolute Error (MAE) as the loss function L I Perform the training of the Transformer Encoder with the following formula: Among them, Class_conf i j represents the confidence score of the i-th image output by the j-th model; Step A6: Fuse the features of J hidden layers through a multi-layer perceptron and output the utility score U i : The multi - layer perceptron includes two fully - connected layers FC1, FC2 and a LeakyReLU activation function. Step A7: Through the loss function L a Perform model training, and the calculation method is as follows: L a = |Score i - U i |, Score i is the utility score of the i-th image in the UDCUI dataset; Step A8: Design a common feature extraction module based on the generative adversarial network GAN. During the training process, define the output of the Transformer Encoder as the input feature where i and j represent the i-th image and the j-th model respectively, the type label of which is 1, and the generator generates features with the type label of 0; the generator consists of two modules, FC'1 and FC'2, each module contains a fully connected layer and a LeakyReLU activation function. First, the feature generates the preliminary common feature through the generator Step A9: A discriminator is composed of three modules, FC3, FC4, and FC5. FC3 contains two fully connected layers, the LeakyReLU, and the Sigmoid activation function; the features After passing through the FC3 module, the output of each neuron of the features is mapped to the interval from 0 to 1: Among them are the features output by FC3; Step A10: Output through the FC4 and FC5 modules of the class probability distribution and the type probability distribution where FC4 is a multi-classification task, including a fully connected layer and a Softmax activation function; FC5 is a binary classification task, including a fully connected layer and a Sigmoid activation function: Step A11: Define the loss function during the training process as: where L S and L V represent the log-likelihoods of the correct type label and the correct class label respectively, and the final loss function L is obtained from this T Formula: Among them, the training objectives of the generator and the discriminator are to minimize L S , L V and maximize L S , L V . Through iterative training, the generator extracts the features that the discriminator cannot correctly classify as the common feature h i ′ j ; Step A12: The common feature h' generated by using the principal component analysis (PCA) transformation generator i j is used for data alignment to remove the specific data in the common feature; for h i ′ 1 to h i ′ J the J features are concatenated to obtain the input f of PCA i , and f i is de-meaned to obtain The mathematical expression of this process is as follows: f i = Concat(h i ′ 1 , h i ′ 2 ... h i ′ J ), wherein is the mean vector of f i ; Step A13, calculate covariance matrix Cov of: where D is the size of the first dimension of and is the transposed matrix of ; perform eigenvalue decomposition on Cov: Cov·V=V·Λ, where V is a matrix composed of eigenvectors, Λ is a diagonal matrix with eigenvalues on the diagonal, and each column eigenvector in V corresponds to the eigenvalue in Λ; arrange the eigenvalues from largest to smallest, and select the eigenvectors corresponding to the top k eigenvalues to form the principal component projection matrix P(v1, v2…v k ); Step A14, the matrix after mean removal Project it into the principal component projection matrix P to obtain the aligned feature Y i : Aligned feature Y i Quality prediction is performed through FC6, which contains two fully connected layers and an activation function LeakyReLU, to obtain the utility score U after common feature extraction i ′: U i ′ = FC6(Y i ), Step A15, define the loss function L during training as follows: f as follows: L f = |Score i - U i '|, Score i It is the utility score of the i-th image in the UDCUI dataset.

7. An image reduction method for underwater target detection according to claim 6, characterized in that Take the image utility evaluation module as the teacher network of the image part, and ResNet18 as the student network to lightweight the image utility evaluation module. Specifically as follows: Step A16: Input image I i Undergoes shallow feature extraction through a convolution Conv1, and then obtains shallow feature F through max pooling Maxpool e , and shallow feature F e Successively undergoes enhanced feature extraction through M Block layers to obtain deep feature F s : F e = Maxpool(Conv1(I i )) F s = BlockM(...(Block2(Block1(F e )))), Step A17: Through the processing of the average pooling Avgpool and the fully connected layer FC1″ layer, the hidden layer feature l is obtained i , the hidden layer feature l i will be mapped into the image utility score S after passing through the fully connected layer FC′2′ i : l i = FC1″(Avgpool(F s )) S i = FC′2′(l i ), Step A18. During training, the lightweight image utility evaluation module constructs a loss function I based on MAE total , which consists of two parts, including feature loss I loss and utility score loss I s . The calculation process is as follows: I total = I loss + I s , I loss ＝|l-l i |, I s = |Score i - S i |, Among them, l is the feature of the hidden layer of the teacher network, that is, the aligned feature Y i 。 8. An image reduction method for underwater target detection according to claim 7, characterized in that, Construct a feature utility evaluation module according to the fish detection model selected in S34 and complete the training using the UDCUI dataset. Specifically including the following steps: Step B1: Modify the selected fish detection model so that the input of the fish detection model can match the dimension of the features in the UDCUI dataset. Step B2: Input the features in the UDCUI dataset into the fish detection model obtained in Step B1 one by one, and use the output features as the migrated utility features F i = [f i,1 ; f i,2 ; …; f i,n , where f i,n represents the feature of a certain scale in F i and n represents the number of different scales of the features. This process is as follows: F i = Detection(f i ), where f i is the feature corresponding to the i-th image in the input UDCUI dataset, and Detection represents the fish detection model obtained in step B1; Step B3: Upsample the low-scale features in the utility feature F i to the intermediate scale, downsample the high-scale features to the intermediate scale, and do not perform additional processing on the intermediate-scale features. Concatenate multiple features of the same scale in the channel dimension to obtain a feature group: P i = [f' i,1 ; f' i,2 …; f' i,n , For the feature group P i Flatten it in the width and height dimensions to obtain a two-dimensional matrix M i ; Step B4: Record the position data of each eigenvector in the two-dimensional matrix in the form of position encoding, that is, add a learnable position matrix POS to matrix M i and add vector f i at the initial end of matrix M r , which is expressed in the following form: M′ i = [f i,r ; f′ i,1 ; f′ i,2 …; f′ i,n + POS, Step B5. The fish detection model obtained in Step B1 is provided with a corresponding Transformer Encoder, and input M i ′ into the Transformer Encoder. The mathematical expression of M i ′ through the Transformer Encoder process is as follows: f″ i,l = MSA(LN(f″′ i,l-1 )) + f″′ i,l-1 , l = 1…n, f′ i,l = MLP(LN(f″ i,l )) + f″ i,l , where f″′ i,l-1 represents the vector in M i ′, f′ i,l represents the output of the TransformerEncoder, and f′ i,l obtains the utility score S through the fully connected layer FC i : S i = FC(f′ i,l ), Step B6: Use the mean absolute error as the loss function L′ I Perform the training of the Transformer Encoder, and the formula is as follows: L′ I = |Class_conf i - S i |, Among them, Class_conf i is the confidence score of the i-th feature in the UDUCI dataset; Step B7. Pass the output feature f′ of the hidden layer i,l through a multi-layer perceptron and output the final utility score U i ″: U i ″ = FC2(LeakyReLU(FC1(f i ′ ,l ))), The multi-layer perceptron includes two fully connected layers FC1, FC2 and a LeakyReLU activation function, and the model is trained through the loss function L′ a The calculation method is as follows: L′ a = |Label i -U i ″|, Label i It is the label of the i-th feature in the UDCUI dataset.

9. An image reduction method for underwater target detection according to claim 8, characterized in that Take the feature utility evaluation module as the teacher network of the feature part, and ResNet18 as the student network to lightweight the feature utility evaluation module, specifically including the following steps: Step B8: Take the input feature f i , and perform shallow feature extraction through the convolutional layer Conv1 to obtain the shallow feature F e '. Subsequently, perform enhanced feature extraction through M Block layers in sequence to obtain the deep feature F s ': F e ′ = Conv1(f i ), F s ′ = BlockM(...(Block2(Block1(F e ′)))), Step B9: After being processed by the average pooling layer and the fully connected layer FC1″, the hidden layer feature l is obtained f , and the hidden layer feature will be mapped into the feature utility score S after passing through the fully connected layer FC′2′ f ; l f = FC1″(Avgpool(F s ′)), S f = FC′2′(l f ), Step B10. During training, the lightweight feature utility evaluation module constructs a loss function F based on the mean absolute error. total It consists of two parts, including the feature loss F loss and the utility score loss F s . The calculation process is as follows: F total = F loss + F s , F loss ＝|l T -l f |, F s = |Label i - S f |, where l T is the output f′ of the multi-layer perceptron in the feature utility evaluation module i,l ; Label i is the label of the i-th feature in the UDCUI dataset.

10. An image reduction system for underwater target detection, characterized in that, It includes a processor, a memory, and a computer program stored on the memory. When the processor executes the computer program, it specifically executes the steps in the image reduction method for underwater target detection as described in any one of claims 1 - 9.

Images