A method for detecting marine benthic organisms based on deep discriminative representation learning under long-tail distribution

Abstract

The present invention discloses a method for detecting benthic organisms based on deep discriminant representation learning under a long-tail distribution, comprising the following steps: constructing a dataset based on images of benthic organisms containing any one or more of sea urchins, scallops, starfish, and sea cucumbers; dividing the dataset into a training set and a test set; constructing a benthic organism detection model based on deep discriminant representation learning under a long-tail distribution; training the benthic organism detection model based on deep discriminant representation learning under a long-tail distribution based on the training set data to obtain a trained benthic organism detection model based on deep discriminant representation learning under a long-tail distribution; inputting the test set data into the trained benthic organism detection model based on deep discriminant representation learning under a long-tail distribution to detect and identify sea urchins, scallops, starfish, and sea cucumbers in the images. The method improves the diversity of the tail-class embedded features of benthic organisms, thereby enhancing the detection accuracy of benthic organisms under a long-tail distribution.

Description

Technical Field

[0001] The present invention belongs to the field of fully automated products and relates to a method for detecting marine benthic organisms based on deep discriminant representation learning under long-tail distribution. Background Art

[0002] Affected by the economic benefits of marine benthic organisms and the way of sample collection, the existing marine benthic organisms dataset [1]-[3] The phenomenon of long-tail distribution is presented in different degrees, where the categories with a large number are called head categories, while the categories with a small number are called tail categories. The detection network based on deep learning can often fully exploit the features of the head category, but it is difficult to effectively learn the features of the tail category, which makes the detector able to detect the head category with high precision, but difficult to accurately detect the tail category. [4]-[6] .

[0003] Methods to improve the accuracy of marine benthic organism detection under long-tail distribution mainly include resampling [7] , data synthesis[6], reweighting [8] and metric learning [9]-

[11] Resampling techniques include oversampling and undersampling. Oversampling involves repeated sampling of rare categories to increase the number of samples, while undersampling involves random sampling of common categories to reduce the number of samples. Data synthesis involves generating images similar to rare categories, while reweighting involves assigning weights to each category based on the number of samples. Due to their efficiency and flexibility, data synthesis and reweighting strategies are more widely used than resampling techniques.

[0004] In order to achieve high-precision detection of marine benthic organisms under long-tail distribution, the gradient field, divergence and coefficient matrix are calculated in sequence.

[12] Poisson fusion technology is used to expand the categories with fewer samples in the marine benthic dataset, thereby improving the detection performance of the few-sample category. In order to increase the number of samples of sea cucumbers and sea urchins, the literature

[13] A generative adversarial network based on Poisson fusion technology is proposed. In the Poisson fusion stage, the targets cropped from the marine benthic organisms dataset are fused with arbitrary seabed backgrounds to increase the number of tail class samples. In the learning stage, U-Net is used to generate more realistic underwater images, effectively improving the accuracy of marine benthic organism detection under long-tail distribution. By comprehensively considering the color, texture and contrast characteristics of the real underwater environment, the literature

[14] We carefully selected images containing only minority-sample classes, constructed an unpaired training dataset, and used an unsupervised cyclic generative adversarial network to keep the number of samples in each category basically consistent, thereby improving the detection accuracy of minority-sample classes.

[0005] Reweighting is another important way to improve the high-precision detection of marine benthic organisms under long-tail distribution

[15] .literature [8] Classification weights are designed based on the proportion of each category, and distance weights are constructed by considering the Euclidean distance between the centers of the predicted and true boxes. This gives higher weights to categories with fewer samples and closer center distances. Results show that combining classification and distance weights can improve detection accuracy by 2.4%.

[0006] In recent years, a large number of experts and scholars have proposed metric learning strategies to solve the long-tail distribution problem by constructing discriminative embedding features. [9] The center loss function is constructed. By introducing the multiplicative angle margin into the Softmax function, the reference

[10] designed the A-Softmax loss to achieve quantitative control of the decision boundary. By applying the normalization technology to the embedded features and weight vectors, the reference

[11] The additive cosine margin was introduced and the Large Margin Cosine Loss (LMCL) technique was developed to help construct discriminative features.

[0007] Through the review and analysis of existing work, the following deficiencies still exist in the fuzzy marine benthic organism detection:

[0008] (1) Regarding sample synthesis methods: Poisson fusion technology ignores the overall consistency of the fused image in terms of contrast, tone and texture, and is overly sensitive to pixels at the boundary between the foreground target and the background image; using generative adversarial networks to achieve data synthesis has problems of training instability and mode collapse; sample synthesis methods such as Cutout and Mixup lack target perception.

[0009] (2) Regarding resampling and reweighting methods: In resampling technology, undersampling the head class easily leads to feature information loss, while oversampling the tail class easily leads to overfitting problems. The reweighting strategy loses the accuracy of the head class to a certain extent.

[0010] (3) Regarding metric learning: Existing center loss only considers the difference between the center of the embedded feature and samples of the same type, ignoring the inter-class dispersion of the embedded features. This makes it difficult to simultaneously improve the compactness of intra-class samples and the dispersion of inter-class samples. The LMCL method cannot maintain a consistent linear decision boundary, which reduces the discriminability of the embedded features. Summary of the Invention

[0011] In order to solve the above problems, the technical solution adopted by the present invention is: a method for detecting marine benthic organisms based on deep discriminant representation learning under long-tail distribution, comprising the following steps:

[0012] Construct a dataset based on images of one or more marine benthic organisms including sea urchins, scallops, starfish, and sea cucumbers;

[0013] Divide the dataset into training and testing sets;

[0014] Construct a marine benthic organism detection model based on deep discriminant representation learning under long-tail distribution;

[0015] Based on the training set data, the marine benthic organism detection model based on deep discriminant representation learning under long-tail distribution is trained to obtain the trained marine benthic organism detection model based on deep discriminant representation learning under long-tail distribution;

[0016] The test set data is input into the marine benthic organism detection model based on deep discriminant representation learning under the trained long-tail distribution to realize the detection and identification of marine benthic organisms such as sea urchins, scallops, starfish and sea cucumbers.

[0017] Furthermore, it is characterized in that: the marine benthic organism detection model based on deep discriminant representation learning under the long-tail distribution includes:

[0018] Backbone network: extracts the informative feature maps contained in the image;

[0019] A classification loss module based on discriminative additional angle margin: Based on the informative feature map extracted by the backbone network and combined with the regional alignment mechanism, a bilinear interpolation method is used to construct the characteristic values of the sampling points in the region of interest of marine benthic organisms. The average aggregation strategy is used to obtain the characteristics of marine benthic organisms in a single interval. The embedded features of marine benthic organisms are constructed with the help of dimensionality conversion operations. The angle between the embedded features and the corresponding class weight is calculated. By increasing the angle margin in the cosine space, the discriminability of the embedded features of marine benthic organisms is improved.

[0020] Uneven embedding feature distribution loss module: A sliding average strategy is used to update the embedding feature centers of various types of marine benthic organisms, and the angles between the embedded features and their corresponding centers are solved in turn to calculate the differences in embedding feature distribution between the tail class and other head classes of marine benthic organisms.

[0021] Furthermore, the classification loss module based on the discriminative additional angle margin includes:

[0022] Informative feature map construction submodule: used to construct the informative feature map extracted by the backbone network;

[0023] A region of interest feature value estimation submodule is configured to estimate the feature value of the region of interest based on the informative feature map transmitted by the informative feature map construction submodule;

[0024] Discriminative additional angle margin classification loss submodule: a loss function is designed based on the eigenvalues of the region of interest estimated by the eigenvalue estimation submodule of the region of interest.

[0025] Furthermore, the process of constructing the informative feature map extracted by the backbone network is as follows:

[0026] The informative feature map extracted by the backbone network is expressed as:

[0027]

[0028] Among them, E(·; ·) is the feature mapping function in the detection network, Represents the height h r , width is w r , batch size is n b and the number of channels is n rc The original input sample, θ e represents the parameters to be optimized, Represents the feature map, h and w are the height and width of the feature map, n c is the number of embedding feature layers.

[0029] Using the batch normalization method, it is expressed as:

[0030]

[0031] Where b = 1, 2, L, n b and l=1,2,L,n c represents the sample index and feature layer index, γ and β represent the scale and offset coefficients respectively, ∈ is a constant, The coordinates of the lth feature layer in the bth sample are (x m ,y n ) at the batch normalized feature value, μ l and σ l are the mean and standard deviation of the features indexed at the lth layer, expressed as:

[0032]

[0033] in, and Represents n c The mean and standard deviation of each channel;

[0034] Taking the b-th sample as an example to illustrate the construction process of discriminative features, in order to increase the nonlinear mapping capability, Leaky ReLU is used as the activation function, which is expressed as:

[0035]

[0036] Among them, 0<κ<1 represents the slope of the activation function, H l (x m ,y n ) represents the coordinates of the lth feature layer (x m ,y n ) is the activation eigenvalue at .

[0037] Furthermore, the process of estimating the feature value of the region of interest based on the informative feature map transmitted by the informative feature map construction submodule is as follows:

[0038] The coordinates of the feature points in the region of interest are expressed as:

[0039]

[0040] in, is the coordinate of the upper left corner of the region of interest, w bin and h bin Represents the width and height of a single interval, g=1,L,n p and k=1,L,n q are the column and row dimension indices of the interval, u=1,L,n col and v=1,L,n row Represents the column and row dimension indices of the sampling points in a single interval, (x g,u ,y k,v ) are the coordinates of the sampling points;

[0041] By means of bilinear interpolation, for the i-th instance in the l-th feature layer, the coordinate position is (x g,u ,y n ) at the eigenvalue Expressed as:

[0042]

[0043] Among them, (x m ,y n ) and (x m+1 ,y n ) is the closest sampling point (x g,u ,y k,v )'s upper left and upper right integer coordinates;

[0044] Similarly, the coordinates (x g,u ,y n+1 ) at the eigenvalue Expressed as:

[0045]

[0046] Among them, (x m ,yn+1 ) and (x m+1 ,y n+1 ) is the closest to (x g,u ,y k,v )'s lower left and lower right corner integer coordinate positions;

[0047] Furthermore, the coordinates are (x g,u ,y k,v ) at the eigenvalue Expressed as:

[0048]

[0049] After training for m batches, for the i-th instance in the l-th feature layer, the feature value at interval index (g, k) Expressed as:

[0050]

[0051] in, Represents the embedding feature of the i-th marine benthic organism instance at the l-th layer. Considering all feature layers, the embedding feature f of the i-th marine benthic organism instance i m Expressed as:

[0052]

[0053] in, represents the embedded features obtained by the dimension transformation operation T(·), d = n q ×n p ×n c , Represents n c The embedded features of channels.

[0054] Furthermore, the process of designing a loss function based on the eigenvalue of the region of interest estimated by the eigenvalue estimation submodule of the region of interest is as follows:

[0055] Normalized embedding features and weights are expressed as:

[0056]

[0057]

[0058] in, yes The jth column of non-normalized weights, j = 1, L, n tc Represents the category index, n tc represents the total number of categories, ‖·‖2 represents the 2-norm, and Represent the normalized embedding features and normalized weights respectively;

[0059] The i-th embedding feature and the weight of class j The angle between them is expressed as:

[0060]

[0061] in,

[0062] Discriminative additional angle margin classification loss function l DAAMCL Expressed as:

[0063]

[0064] Where s is the scale parameter, δ represents the additional angle margin, Represents embedded features and corresponding category weights The angle between them, n inst Represents the number of sample instances in this batch.

[0065] Furthermore, the distribution loss module of the uneven embedding feature includes:

[0066] The embedding feature center construction submodule is used to transfer the distribution information of the embedded features of marine benthic organisms from the implicit head class to the tail class, propose an uneven embedding feature distribution loss, enhance the diversity of the embedded features of the tail class of marine benthic organisms, and construct the embedded features of each category separately;

[0067] Intra-class embedding feature distribution variance calculation submodule: used to calculate the embedding feature distribution variance of each category of embedding features constructed by the embedding feature center construction submodule;

[0068] Uneven embedding feature distribution loss submodule: used for uneven embedding feature distribution loss function of all categories based on the embedding feature distribution variance calculated by the intra-class embedding feature distribution variance calculation submodule.

[0069] Furthermore, by migrating the distribution information of the embedded features of marine benthic organisms from the implicit head class to the tail class, an uneven embedded feature distribution loss is proposed to enhance the diversity of the embedded features of the tail class of marine benthic organisms. The process of constructing the embedded features of each category is as follows:

[0070] Construct the embedding features of each category and calculate the corresponding embedding feature center, which is expressed as:

[0071]

[0072] Among them, Ω∈{ec,sc,st,ho}, ec, sc, st and ho represent the categories of sea urchin, scallop, starfish and sea cucumber respectively, represents the i-th embedded feature in the corresponding category after training m batches, represents the center of the embedded feature, Represents the number of samples of the corresponding category after training m batches;

[0073] Adopting the sliding average mechanism, fully utilizing the embedding feature center of the m-1th batch To update the embedding feature center of the mth batch Expressed as:

[0074]

[0075] in, λ represents the class center update rate.

[0076] Furthermore, the process of constructing the embedding features of each category and calculating the embedding feature distribution variance of the embedding feature center construction submodule is as follows:

[0077] The angle between the embedded feature and its corresponding category center is expressed as:

[0078]

[0079] in, is the i-th embedding feature after training m batches Its corresponding center Angle.

[0080] Average values of angular distribution for different categories Expressed as:

[0081]

[0082] Subsequently, the intra-class embedding feature distribution variance is expressed as:

[0083]

[0084] in, and Represents the variance of the intra-class embedding feature distribution of sea urchins, scallops, starfish, and sea cucumbers after training for m batches.

[0085] Furthermore, based on the embedding feature distribution variance calculated by the intra-class embedding feature distribution variance calculation submodule, the process of the uneven embedding feature distribution loss function for all categories is as follows:

[0086] The uneven embedding feature distribution loss is expressed as:

[0087]

[0088] in, and Representatives from the set The first largest intra-class embedding feature distribution variance, the second largest intra-class embedding feature distribution variance, the third largest intra-class embedding feature distribution variance and the fourth largest intra-class embedding feature distribution variance, n 1st 、n 2nd 、n 3rd and n 4th is the corresponding number of instances, and Represents the uneven embedding feature distribution loss for different categories.

[0089] Finally, the uneven embedding feature distribution loss l for all categories UEFDL Expressed as:

[0090]

[0091] The present invention provides a method for detecting benthic organisms based on deep discriminant representation learning under a long-tail distribution, which has the following advantages:

[0092] 1. Using regional alignment technology, the spatial correspondence between the region of interest of marine benthic organisms and the embedded features can be established, avoiding the mapping offset problem caused by quantization operations and ensuring the integrity of the embedded features of benthic organisms;

[0093] 2. By calculating the angle between the embedding features of marine benthic organisms and the corresponding class weights, an additional angular margin is added in the cosine space, which improves the discriminability of the embedding features of marine benthic organisms and enhances the intra-class compactness and inter-class differentiation.

[0094] 3. Without manually specifying the head and tail classes, the embedded feature distribution information of marine benthic organisms can be adaptively transferred from the implicit head class to the tail class, thereby improving the diversity of the embedded features of the tail class of marine benthic organisms and thus enhancing the detection accuracy of marine benthic organisms under long-tail distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0096] Figure 1 A flowchart of the application method;

[0097] Figure 2 It is a sampling point eigenvalue calculation mechanism with a bilinear interpolation strategy;

[0098] Figure 3 An additional angle margin classification loss is added for discriminability;

[0099] Figure 4 The loss is the uneven embedding feature distribution;

[0100] Figure 5 Visualization of the embedding feature distribution for different categories; (a) is the feature distribution of the 1st iteration, (b) is the feature distribution of the 20th iteration, (c) is the feature distribution of the 40th iteration, (d) is the feature distribution of the 60th iteration, (e) is the feature distribution of the 80th iteration, and (f) is the feature distribution of the 100th iteration;

[0101] Figure 6 The angular distribution of different categories: (a) angular distribution of sea urchins, (b) angular distribution of scallops, (c) angular distribution of starfish, and (d) angular distribution of sea cucumbers.

[0102] Figure 7 Performance comparison of the precision-recall curves of different categories, including (a) the precision-recall curve of sea urchins, (b) the precision-recall curve of scallops, (c) the precision-recall curve of starfish, and (d) the precision-recall curve of sea cucumbers;

[0103] Figure 8 Qualitative comparison of detection performance in typical scenarios, where (a) is the dark light situation, (b) is the bright light situation, (c) is the small scale situation, and (d) is the coexistence of multiple categories of targets. DETAILED DESCRIPTION

[0104] It should be noted that, unless there is any conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0105] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is in no way intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0106] Figure 1 A flowchart of the application method;

[0107] A method for detecting marine benthic organisms based on deep discriminant representation learning under a long-tail distribution includes the following steps:

[0108] S1: Construct a dataset based on images of any one or more marine benthic organisms including sea urchins, scallops, starfish, and sea cucumbers;

[0109] S2: Divide the dataset into training set and test set;

[0110] S3: Construct a marine benthic organism detection model based on deep discriminant representation learning under long-tail distribution;

[0111] S4: Based on the training set data, the marine benthic organism detection model based on deep discriminant representation learning under the long-tail distribution is trained to obtain the trained marine benthic organism detection model based on deep discriminant representation learning under the long-tail distribution;

[0112] S5: Input the test set data into the marine benthic organism detection model based on deep discriminant representation learning under the trained long-tail distribution to realize the detection and identification of marine benthic organisms such as sea urchins, scallops, starfish and sea cucumbers.

[0113] Steps S1 / S2 / S3 / S4 / S5 are performed sequentially;

[0114] Furthermore: the marine benthic organism detection model based on deep discriminant representation learning under the long-tail distribution includes:

[0115] Backbone network: extracts the informative feature maps contained in the image;

[0116] A classification loss module based on discriminative additional angle margin: Based on the informative feature map extracted by the backbone network and combined with the regional alignment mechanism, a bilinear interpolation method is used to construct the characteristic values of the sampling points in the region of interest of marine benthic organisms. The average aggregation strategy is used to obtain the characteristics of marine benthic organisms in a single interval. The embedded features of marine benthic organisms are constructed with the help of dimensionality conversion operations. The angle between the embedded features and the corresponding class weight is calculated. By increasing the angle margin in the cosine space, the discriminability of the embedded features of marine benthic organisms is improved.

[0117] Uneven embedding feature distribution loss module: A sliding average strategy is used to update the embedding feature centers of various types of marine benthic organisms, and the angles between the embedded features and their corresponding centers are solved in turn to calculate the differences in embedding feature distribution between the tail class and other head classes of marine benthic organisms.

[0118] The two loss modules collaboratively optimize the entire detection and recognition network model;

[0119] Furthermore: the classification loss module based on the discriminative additional angle margin includes:

[0120] Informative feature map construction submodule: used to construct the informative feature map extracted by the backbone network;

[0121] A region of interest feature value estimation submodule is configured to estimate the feature value of the region of interest based on the informative feature map transmitted by the informative feature map construction submodule;

[0122] Discriminative additional angle margin classification loss submodule: a loss function is designed based on the eigenvalues of the region of interest estimated by the eigenvalue estimation submodule of the region of interest.

[0123] Furthermore, the process of constructing the informative feature map extracted by the backbone network is as follows:

[0124] The informative feature map extracted by the backbone network is expressed as:

[0125]

[0126] Among them, E(·; ·) is the feature mapping function in the detection network, Represents the height h r , width is w r , batch size is n b and the number of channels is n rc The original input sample, θ e represents the parameters to be optimized, Represents the feature map, h and w are the height and width of the feature map, n c is the number of embedding feature layers.

[0127] Using the batch normalization method, it is expressed as:

[0128]

[0129] Where b = 1, 2, L, n b and l=1,2,L,n c represents the sample index and feature layer index, γ and β represent the scale and offset coefficients respectively, ∈ is a constant, The coordinates of the lth feature layer in the bth sample are (x m ,y n ) at the batch normalized feature value, μ l and σ l are the mean and standard deviation of the features indexed at the lth layer, expressed as:

[0130]

[0131] in, and Represents n cThe mean and standard deviation of each channel;

[0132] Taking the b-th sample as an example to illustrate the construction process of discriminative features, in order to increase the nonlinear mapping capability, Leaky ReLU is used as the activation function, which is expressed as:

[0133]

[0134] Among them, 0<κ<1 represents the slope of the activation function, H l (x m ,y n ) represents the coordinates of the lth feature layer (x m ,y n ) is the activation eigenvalue at .

[0135] Figure 2 It is a sampling point eigenvalue calculation mechanism with a bilinear interpolation strategy;

[0136] Furthermore, the process of estimating the feature value of the region of interest based on the informative feature map transmitted by the informative feature map construction submodule is as follows:

[0137] The coordinates of the feature points in the region of interest are expressed as:

[0138]

[0139] in, is the coordinate of the upper left corner of the region of interest, w bin and h bin Represents the width and height of a single interval, g=1,L,n p and k=1,L,n q are the column and row dimension indices of the interval, u=1,L,n col and v=1,L,n row Represents the column and row dimension indices of the sampling points in a single interval, (x g,u ,y k,v ) are the coordinates of the sampling points;

[0140] By means of bilinear interpolation, for the i-th instance in the l-th feature layer, the coordinate position is (x g,u ,y n ) at the eigenvalue Expressed as:

[0141]

[0142] Among them, (x m ,y n ) and (x m+1 ,y n ) is the closest sampling point (xg,u ,y k,v )'s upper left and upper right integer coordinates;

[0143] Similarly, the coordinates (x g,u ,y n+1 ) at the eigenvalue Expressed as:

[0144]

[0145] Among them, (x m ,y n+1 ) and (x m+1 ,y n+1 ) is the closest to (x g,u ,y k,v )'s lower left and lower right corner integer coordinate positions;

[0146] Furthermore, the coordinates are (x g,u ,y k,v ) at the eigenvalue Expressed as:

[0147]

[0148] After training for m batches, for the i-th instance in the l-th feature layer, the feature value at interval index (g, k) Expressed as:

[0149]

[0150] in, Represents the embedding feature of the i-th marine benthic organism instance at the l-th layer. Considering all feature layers, the embedding feature f of the i-th marine benthic organism instance i m Expressed as:

[0151]

[0152] in, represents the embedded features obtained by the dimension transformation operation T(·), d = n q ×n p ×n c , Represents n c The embedded features of channels.

[0153] Figure 3 An additional angle margin classification loss is added for discriminability;

[0154] Furthermore, the process of designing a loss function based on the eigenvalue of the region of interest estimated by the eigenvalue estimation submodule of the region of interest is as follows:

[0155] Normalized embedding features and weights are expressed as:

[0156]

[0157] in, yes The jth column of non-normalized weights, j = 1, L, n tc Represents the category index, n tc represents the total number of categories, ‖·‖2 represents the 2-norm, and Represent the normalized embedding features and normalized weights respectively;

[0158] The i-th embedding feature and the weight of class j The angle between them is expressed as:

[0159]

[0160] in,

[0161] Discriminative additional angle margin classification loss function l DAAMCL Expressed as:

[0162]

[0163] Where s is the scale parameter, δ represents the additional angle margin, Represents embedded features and corresponding category weights The angle between them, n inst Represents the number of sample instances in this batch.

[0164] Furthermore: the distribution loss module of the uneven embedding feature includes:

[0165] The embedding feature center construction submodule is used to transfer the distribution information of the embedded features of marine benthic organisms from the implicit head class to the tail class, propose an uneven embedding feature distribution loss, enhance the diversity of the embedded features of the tail class of marine benthic organisms, and construct the embedded features of each category separately;

[0166] Intra-class embedding feature distribution variance calculation submodule: used to calculate the embedding feature distribution variance of each category of embedding features constructed by the embedding feature center construction submodule;

[0167] Uneven embedding feature distribution loss submodule: used for uneven embedding feature distribution loss function of all categories based on the embedding feature distribution variance calculated by the intra-class embedding feature distribution variance calculation submodule.

[0168] Furthermore, by migrating the embedded feature distribution information of marine benthic organisms from the implicit head class to the tail class, an uneven embedded feature distribution loss is proposed to enhance the diversity of embedded features of the tail class of marine benthic organisms. The process of constructing the embedded features of each category is as follows:

[0169] Construct the embedding features of each category and calculate the corresponding embedding feature center, which is expressed as:

[0170]

[0171] Among them, Ω∈{ec,sc,st,ho}, ec, sc, st and ho represent the categories of sea urchin, scallop, starfish and sea cucumber respectively, represents the i-th embedded feature in the corresponding category after training m batches, represents the center of the embedded feature, Represents the number of samples of the corresponding category after training m batches.

[0172] Adopting the sliding average mechanism, fully utilizing the embedding feature center of the m-1th batch To update the embedding feature center of the mth batch Expressed as:

[0173]

[0174] in, λ represents the class center update rate.

[0175] Furthermore, the process of constructing the embedding features of each category and calculating the embedding feature distribution variance of the embedding feature center construction submodule is as follows:

[0176] The angle between the embedded feature and its corresponding category center is expressed as:

[0177]

[0178] in, is the i-th embedding feature after training m batches Its corresponding center Angle.

[0179] Average values of angular distribution for different categories Expressed as:

[0180]

[0181] Subsequently, the intra-class embedding feature distribution variance is expressed as:

[0182]

[0183] in, and Represents the variance of the intra-class embedding feature distribution of sea urchins, scallops, starfish, and sea cucumbers after training for m batches.

[0184] Figure 3 The loss is the uneven embedding feature distribution;

[0185] Furthermore, based on the embedding feature distribution variance calculated by the intra-class embedding feature distribution variance calculation submodule, the process of the uneven embedding feature distribution loss function for all categories is as follows:

[0186] The uneven embedding feature distribution loss is expressed as:

[0187]

[0188] in, and Representatives come from sets The first largest intra-class embedding feature distribution variance, the second largest intra-class embedding feature distribution variance, the third largest intra-class embedding feature distribution variance and the fourth largest intra-class embedding feature distribution variance, n 1st 、n 2nd 、n 3rd and n 4th is the corresponding number of instances, and Represents the uneven embedding feature distribution loss for different categories.

[0189] Finally, the uneven embedding feature distribution loss l for all categories UEFDL Expressed as:

[0190]

[0191] Example 1:

[0192] 1 Dataset and Experimental Configuration Dataset and Experimental Configuration

[0193] To verify the effectiveness and superiority of the DDRL method, a comprehensive comparative experiment was conducted on the URPC2020 dataset and the DUO dataset, which contain 5543 and 7782 images respectively. For the URPC2020 dataset, the training set contains 4434 images and the test set contains 1109 images; for the DUO dataset, the training set contains 6671 images and the test set contains 1111 images. The corresponding experimental configuration parameters are shown in Table 1. The constant ∈ = 10 is set based on experience. -5and scale parameter s = 64. AP, mAP@.5, mAP@.75, and mAP@[.5,.95] were used as detection performance evaluation metrics. A five-fold cross-validation strategy was used, and the average of the five test results was taken as the detection accuracy of different models. Inference speed was obtained using videos with a resolution of 1280×720.

[0194] Table 1 Experimental configuration parameters

[0195]

[0196] 2 Model parameter selection

[0197] In order to obtain the optimal number of embedding feature layers n c 、Number of intervals (n p and n q ), number of sampling points (n col and n row ), sampling strategy, angle margin δ and class center update rate λ, and carried out corresponding performance comparison experiments. In the present invention, set n p and n q Keep consistent, col and n row Keeping the same, the comparison results are shown in Table 2-Table 6.

[0198] Table 2 Performance comparison of different embedding feature layers

[0199]

[0200] Table 3 Performance comparison of different numbers of intervals and sampling points

[0201]

[0202] Table 4 Performance comparison of different sampling strategies

[0203]

[0204] Table 5 Performance comparison of different angle margins

[0205]

[0206] Table 6 Performance comparison of different center update rates

[0207]

[0208] It can be clearly seen from Table 2 that with the increase of the number of embedding feature layers, the detection accuracy of different categories gradually improves, which shows that more discriminative embedding features of marine benthic organisms are constructed. c =256 to n c=512, the detection accuracy improved is quite limited. In order to save computing resources, in subsequent experiments, n is set to c =256.

[0209] In addition, regarding the determination of the optimal number of intervals (n p and n q ) and the number of sampling points (n col and n row ), Table 3 clearly shows that detection accuracy is affected by the number of intervals and sampling points, and it is difficult to achieve optimal detection accuracy for all four categories using the same set of parameters. Taking all detection accuracy into consideration, we ultimately chose 4 intervals and 5 sampling points.

[0210] Next, regarding the optimal sampling strategy, Table 4 shows that the detection accuracy achieved by the RoIPool and RoIWarp methods is significantly lower than that achieved by the region alignment method. This is primarily due to the quantization used by the RoIPool and RoIWarp methods, which results in a mapping offset between the region of interest and the constructed embedded features. In contrast, the region alignment method, using bilinear interpolation techniques, avoids the quantization process, allowing for pixel-by-pixel spatial correspondences. Furthermore, the average aggregation strategy achieves higher detection accuracy.

[0211] Furthermore, regarding the determination of the optimal angle margin δ, Table 5 shows that detection accuracy increases with increasing angle margin δ, indicating that a more discriminative embedding feature is constructed. Although the sea urchin class achieves the highest detection accuracy when δ = 0.55, in order to balance the performance of the mAP@.5, mAP@.75, and mAP@[.5,.95] indicators, δ = 0.45 is set in subsequent experiments.

[0212] To determine the optimal class center update rate λ, we conducted performance comparison experiments, with the results shown in Table 6. Overall, as the class center update rate λ increases, the detection accuracy of all classes gradually improves. However, when λ>0.3, the detection accuracy gradually decreases, and we ultimately set λ=0.3.

[0213] 3 Embedded Feature Visualization

[0214] The t-Distributed Stochastic Neighbor Embedding (t-SNE) technique is used to present the distribution of embedding features of different categories. The corresponding results are as follows: Figure 5 As shown, Figure 5Visualization of the distribution of embedded features for different categories; (a) is the feature distribution of the 1st iteration, (b) is the feature distribution of the 20th iteration, (c) is the feature distribution of the 40th iteration, (d) is the feature distribution of the 60th iteration, (e) is the feature distribution of the 80th iteration, and (f) is the feature distribution of the 100th iteration. It is difficult to distinguish the embedded features of different categories from (a) and (b), and there is no clear boundary between the classes. As the number of iterations increases, DDRL produces more discriminative embedded features, as shown in (c) and (d). As shown in (e) and (f), with the help of the proposed DAAMCL module, DDRL can derive more obvious boundaries between different classes. In particular, for the scallop, starfish, and sea cucumber categories, the spatial range derived at the 100th iteration is larger than that at the 20th iteration, which clearly shows that with the help of the proposed UEFDL module, the tail class can gradually learn the intra-class diversity from the implicit head class.

[0215] also, Figure 6 The angular distribution of different categories: (a) angular distribution of sea urchins, (b) angular distribution of scallops, (c) angular distribution of starfish, and (d) angular distribution of sea cucumbers.

[0216] The initial and final angular distributions of the embedded features and their corresponding class centers are shown. As can be seen, the initial and final angular distributions exhibit a Gaussian shape. Specifically, the initial angular distribution centers for sea urchins, scallops, starfish, and sea cucumbers are approximately 40°, 50°, 43°, and 40°, respectively. In this case, the embedded features lack discriminative properties. Combining the DAAMCL and UEFDL modules, after multiple iterative training cycles, achieves a more compact intra-class distribution, with final angular distribution centers approximately 20°, 25°, 23°, and 20°. This demonstrates that DDRL derives more discriminative embedded features.

[0217] 4-module ablation experiment

[0218] To fully demonstrate the contributions of the DAAMCL and UEFDL modules to the DDRL approach, we selected YOLOv3 as the baseline model. The corresponding experimental results are summarized in Table 7. As can be seen, the combination of the DAAMCL module achieves superior detection performance compared to the baseline model. The key reason is that the additional angular margin helps construct discriminative embedding features. Furthermore, the combination of the UEFDL module improves the mAP@.5, mAP@.75, and mAP@[.5,.95] metrics, demonstrating that the variance of the embedding feature distribution can be transferred from the head class to the tail class. The designed DDRL, combining the DAAMCL and UEFDL modules, achieves the best detection performance.

[0219] Table 7 Ablation experiments of different modules

[0220]

[0221] Accordingly, Figure 7 The performance comparison of the precision-recall curves for different categories is shown, including (a) the precision-recall curve for sea urchins, (b) the precision-recall curve for scallops, (c) the precision-recall curve for starfish, and (d) the precision-recall curve for sea cucumbers. The precision-recall curves for different categories show that the proposed DDRL method can achieve an optimal balance between precision and recall. In particular, for scallops, starfish, and sea cucumbers, the combination of the DAAMCL module and the UEFDL module can significantly improve detection accuracy. The key reasons include: 1) By increasing the angular margin between the embedded features and the corresponding class weights in cosine space, the designed DAAMCL module improves intra-class compactness and inter-class differentiation; 2) The UEFDL module can adaptively transfer the distribution information of the embedded features of marine benthic organisms from the implicit head class to the tail class, improving the detection accuracy of the tail class while maintaining the detection performance of the head class.

[0222] 5 Quantitative comparison of classic long-tail object detection methods

[0223] To demonstrate the effectiveness and superiority of DDRL, we comprehensively considered reweighting methods (Class-Balanced Softmax, Focal Loss, and LDAM) and data synthesis methods (Cutout, Random Erasing, GridMask, and Poisson GAN). YOLOv3 was selected as the baseline model, and the corresponding comparative experimental results are summarized in Table 8. As can be seen, the Class-Balanced Softmax technique combined with the method achieves higher detection accuracy than the baseline model, based on mAP@.5, mAP@.75, and mAP@[.5,.95] metrics. In addition to the detection accuracy of sea urchins, the focal loss-based detection method also surpasses the baseline model. Although the loss function with known label distribution can provide greater margin for tail classes, the LDAM method still achieves lower detection accuracy than the DDRL method, primarily due to its failure to consider the distribution of embedded features across different classes. Regarding data synthesis methods, the Cutout method, by randomly removing fixed-size regions, facilitates the detection network to comprehensively learn the global semantic information of the objects of interest. In particular, when detecting sea cucumbers, the Random Erasing technique significantly improves detection performance by arbitrarily selecting square regions and replacing them with random values. Although the GridMask method avoids over-deletion and retains continuous regions, its detection accuracy is lower than that of DDRL, except for sea urchins and sea cucumbers. Combining Poisson fusion technology with a generative adversarial network, which increases the number of samples in the tail class, the Poisson GAN achieves the best sea cucumber and mAP@.5 metric values. Based on the mAP@[.5,.95] metric values, the proposed DDRL comprehensively improves the detection performance of marine benthic organisms.

[0224] Table 8 Quantitative comparison of classic long-tail target detection methods on the URPC2020 dataset

[0225]

[0226] Quantitative comparison of 6 specialized marine benthic organism detection methods

[0227] In order to verify the superiority of DDRL, the present invention considers comparing it with a dedicated marine benthic organism detection method, including the Lightweight Underwater Object Detection (LUOD) method, the Hybrid Transformer-based Detection Network (HTDN), TVFRD, XC-SSD, BBROABR and FCSA. The corresponding experimental comparison results are shown in Table 9. It can be seen that since the LUOD method adopts the lightweight backbone network MobileNet v2, it achieves the fastest detection speed (44FPS). Compared with the ResNet-50 feature extraction network, the XC-SSD based on ResNeXt-50 achieves higher detection accuracy. The main reason is that ResNeXt-50 adopts the grouping idea and increases the representation ability of the convolutional neural network. When detecting the sea cucumber category, XC-SSD achieved the best detection accuracy. HTDN uses the Transformer structure to enhance the ability to obtain global and local texture information, and obtains the highest accuracy when detecting sea urchins. Because TVFRD relies on a region proposal network to generate a large number of candidate bounding boxes, its inference speed can only reach 3 FPS. Furthermore, although BBROABR and FCSA enhance detection performance by improving bounding box regression localization and recalibrating relevant feature map weights, respectively, their detection accuracy is lower than that of the proposed DDRL. By developing the DAAMCL and UEFDL modules, the proposed DDRL method transfers the variance of the embedding feature distribution of the head class to the tail class, surpassing other methods in terms of mAP@.5, mAP@.75, and mAP@[.5,.95] metrics.

[0228] Table 9 Quantitative comparison of dedicated marine benthic organism detection methods on the DUO dataset

[0229]

[0230] 7 Quantitative comparison of common target detection methods

[0231] Furthermore, the present invention comprehensively considers two-stage detection methods (Faster R-CNN), single-stage detection methods (SSD, EfficientDet, YOLOv4, YOLOv7 and RetinaNet) and anchor-free detection methods (FCOS and CenterNet), and the corresponding comparison results are summarized in Table 10. It can be seen that due to the lack of prior anchor box information, anchor-free detection methods (i.e., FCOS and CenterNet) have difficulty in achieving superior detection performance. Although YOLOv4 uses a more efficient backbone network CSPDarkNet-53 to extract high-level semantic information, it has difficulty in constructing discriminative embedding features due to the lack of an additional angle margin strategy. YOLOv7 utilizes an extended efficient layer aggregation network, a hierarchical label assignment method from coarse-grained to fine-grained, and a convolutional reparameterization strategy to effectively improve feature extraction and network learning capabilities, achieving the best value in terms of the mAP@.75 indicator. The SSD and EfficientDet methods do not consider transferring the embedded feature distribution information, and therefore their detection accuracy is lower than that of DDRL. Faster R-CNN, using ResNet-50 as the feature extraction network, achieves higher detection accuracy than VGG-16. However, because Faster R-CNN requires a large number of candidate boxes generated by the region proposal network, it can only achieve an inference speed of 5 FPS at a resolution of 1000×600. Combining the DAAMCL and UEFDL modules, the proposed DDRL surpasses other methods in terms of mAP@.5 and mAP@[.5,.95] metrics. The improvement in detection accuracy for sea cucumbers and scallops is greater than that for sea urchins, indicating that the UEFDL module plays a crucial role in transferring the feature distribution information of the head class embedding.

[0232] Table 10 Quantitative comparison of general target detection methods on the URPC2020 dataset

[0233]

[0234] 8 Qualitative comparison of marine benthic organism detection

[0235] To demonstrate the superiority of DDRL in real-world detection, we consider two-stage detection methods (Faster R-CNN), single-stage detection methods (SSD, EfficientDet, YOLOv4, and YOLOv7), and anchor-free detection methods (FCOS). Experiments are conducted in low-light scenes, bright-light scenes, small-scale target scenes, and scenes with coexistence of multiple categories of targets. Figure 8 Qualitative comparison of detection performance in typical scenarios, where (a) is low-light, (b) is bright-light, (c) is small-scale, and (d) is multi-category target coexistence.

[0236] As shown in Figure (a), Faster R-CNN achieves good detection results in low-light scenarios, thanks to the high-quality bounding boxes generated by the region proposal network. Regarding detection performance in bright-light scenarios, as shown in Figure (b), Faster R-CNN and EfficientDet misidentify seagrass as sea cucumbers, while SSD and YOLOv4 fail to correctly detect and identify benthic organisms located in rock crevices. Regarding detection performance in small-scale benthic organisms, as shown in Figure (c), Faster R-CNN, YOLOv7, and DDRL all significantly reduce false detections. Because FCOS does not use prior anchor box information during training, it struggles to accurately detect all benthic organisms. SSD and EfficientDet lack discriminative embedding features and therefore can only detect a subset of benthic organisms. Regarding detection performance in scenarios with multiple classes of objects, as shown in Figure (d), DDRL accurately detects benthic organisms that are both intra-class and inter-class occluded.

[0237] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

[0238] [1]Spampinato C,Palazzo S,Joalland P,et al.Fine-grained objectrecognition in underwater visual data[J].Multimedia Tools and Applications,2016,75:1701-1720.

[0239] [2]Qin H, Li

[0240] [3]MaY,Liu M,ZhangY,etal.Imbalanced underwater acoustic targetrecognition with trigonometric loss and attention mechanism convolutionalnetwork[J].Remote Sensing,2022,14(16):4103.

[0241] [4]Xu S,Zhang M,Song W,et al.A systematic review and analysis of deeplearning-based underwater object detection[J].Neurocomputing,2023,527:204-232.

[0242] [5]NieW,RenJ,ChengG,et al.Tackling long-tail data distributionproblem of deep learning based underwater target recognition system[C].OCEANSConference,San Diego,USA,2021:1-6.

[0243] [6]LiY,SunK,Han Z.Vision technology in underwater:Applications,challenges and perspectives[C].The 4th International Conference on Controland Robotics,Guangzhou,China,2022:369-378.

[0244] [7]PangY,Wu C,Wu H,et al.Over-sampling strategy-based class-imbalanced salient object detection and its application in underwater scene[J].The Visual Computer,2023,39(5):1959-1974.

[0245] [8]Zhang Z,Liu Y,Zhu X,et al.DSE-FCOS:Dilated and SE block-reinforcedFCOS for detection of marine benthos[J].The Visual Computer,2023,39(11):5199-5208.

[0246] [9]Wen Y,Zhang K,Li Z,et al.A discriminative feature learningapproach for deep face recognition[C].European Conference on Computer Vision,Amsterdam,Netherlands,2016:499-515.

[0247]

[10] Liu W,Wen Y,Raj B,et al.Sphereface revived:Unifyinghyperspherical face recognition[J].IEEE Transactions on Pattern Analysis andMachine Intelligence,2022,45(2):2458-2474.

[0248]

[11] Wang H,Wang Y,Zhou Z,et al.Cosface:Large margin cosine loss fordeep face recognition[C].IEEE Conference on Computer Vision and PatternRecognition,Salt Lake City,USA,2018:5265-5274.

[0249]

[12] Wang J,Qi S,Wang C,et al.B-YOLOX-S:A lightweight method forunderwater object detection based on data augmentation and multiscale featurefusion[J].Journal of Marine Science and Engineering,2022,10(11):1764.

[0250]

[13] Liu C,Wang Z,Wang S,et al.A new dataset,Poisson GAN and AquaNetfor underwater object grabbing[J].IEEE Transactions on Circuits and Systemsfor Video Technology,2022,32(5):2831-2844.

[0251]

[14] Chen L,Dong J,Zhou H.Class balanced underwater object detectiondataset generated by class-wise style augmentation[J].arXiv preprint:2101.07959,2021.

[0252]

[15] Buda M,Maki A,Mazurowski M A.A systematic study of the classimbalance problem in convolutional neural networks[J].Neural Networks,2018,106:249-259.

Claims

1. A method for detecting benthic organisms based on deep discriminant representation learning under long-tail distribution, characterized by: The following steps are involved: Construct a dataset based on images of one or more marine benthic organisms including sea urchins, scallops, starfish, and sea cucumbers; Divide the dataset into training and testing sets; Construct a marine benthic organism detection model based on deep discriminant representation learning under long-tail distribution; Based on the training set data, the marine benthic organism detection model based on deep discriminant representation learning under long-tail distribution is trained to obtain the trained marine benthic organism detection model based on deep discriminant representation learning under long-tail distribution; The test set data is input into the trained long-tail distribution-based marine benthic organism detection model based on deep discriminant representation learning to achieve the detection and identification of sea urchins, scallops, starfish and sea cucumbers. The marine benthic organism detection model based on deep discriminant representation learning under the long-tail distribution includes: Backbone network: extracts the informative feature maps contained in the image; A classification loss module based on discriminative additional angle margin: Based on the informative feature map extracted by the backbone network and combined with the regional alignment mechanism, a bilinear interpolation method is used to construct the characteristic values of the sampling points in the region of interest of marine benthic organisms. The average aggregation strategy is used to obtain the characteristics of marine benthic organisms in a single interval. The embedded features of marine benthic organisms are constructed with the help of dimensionality conversion operations. The angle between the embedded features and the corresponding class weight is calculated. By increasing the angle margin in the cosine space, the discriminability of the embedded features of marine benthic organisms is improved. Uneven embedding feature distribution loss module: A sliding average strategy is used to update the embedding feature centers of various types of marine benthic organisms, and the angles between the embedded features and their corresponding centers are solved in turn to calculate the differences in embedding feature distribution between the tail class and other head classes of marine benthic organisms.

2. The method for detecting benthic organisms based on deep discriminant representation learning under a long-tail distribution according to claim 1, characterized in that: The classification loss module based on the discriminative additional angle margin includes: Informative feature map construction submodule: used to construct the informative feature map extracted by the backbone network; A region of interest feature value estimation submodule is configured to estimate the feature value of the region of interest based on the informative feature map transmitted by the informative feature map construction submodule; Discriminative additional angle margin classification loss submodule: a loss function is designed based on the eigenvalues of the region of interest estimated by the eigenvalue estimation submodule of the region of interest.

3. The method for detecting benthic organisms based on deep discriminant representation learning under a long-tail distribution according to claim 1, characterized in that: The process of constructing the informative feature map extracted by the backbone network is as follows: The informative feature map extracted by the backbone network is expressed as: (1) in, is the feature mapping function in the detection network, Representative height is , width is , the batch size is and the number of channels is The original input sample, represents the parameters to be optimized, Represents the feature map, and are the height and width of the feature map, is the number of embedding feature layers; Using the batch normalization method, it is expressed as: (2) in, and Represents the sample index and feature layer index, and represent the scale and offset coefficients respectively, is a constant, Representative In the sample The feature layer coordinates are The batch normalized eigenvalues at , and The indexes are The mean and standard deviation of the layer features are expressed as: (3) in, and Respectively represent The mean and standard deviation of each channel; First The construction process of discriminative features is explained using a sample as an example. In order to increase the nonlinear mapping capability, LeakyReLU is used as the activation function, which is expressed as: (4) in, represents the slope of the activation function, Representative The coordinates in the feature layer are The activation eigenvalue at .

4. The method for detecting benthic organisms based on deep discriminant representation learning under a long-tail distribution according to claim 2, characterized in that: The process of estimating the feature value of the region of interest based on the informative feature map transmitted by the informative feature map construction submodule is as follows: The coordinates of the feature points in the region of interest are expressed as: (5) in, is the coordinate of the upper left corner of the region of interest, and Represents the width and height of a single interval respectively, and are the column and row dimension indices of the interval, and Respectively represent the index of the sampling point in a single interval in the column dimension and row dimension, are the coordinates of the sampling points; By means of bilinear interpolation method, for the In the feature layer For an instance, the coordinate position is The eigenvalue at Expressed as: (6) in, and The closest sampling point The integer coordinates of the upper left and upper right corners; Similarly, the coordinates are The eigenvalue at Expressed as: (7) in, and is closest The integer coordinate positions of the lower left and lower right corners; Furthermore, the coordinates are The eigenvalue at Expressed as: (8) In training After batches, for In the feature layer For an instance, the interval index is The eigenvalue at Expressed as: (9) in, Representative Examples of marine benthic organisms are in The embedding features of the layer, considering all feature layers, Embedded features of marine benthic organisms Expressed as: (10) in, Represents a dimension conversion operation The embedded features obtained are , represent The embedded features of channels.

5. The method for detecting benthic organisms based on deep discriminant representation learning under a long-tail distribution according to claim 2, characterized in that: The process of designing the loss function based on the eigenvalue of the region of interest estimated by the eigenvalue estimation submodule of the region of interest is as follows: Normalized embedding features and weights are expressed as: (11) (12) in, yes Middle Column non-normalized weights, Represents the category index, represents the total number of categories, represents the 2-norm, and Represent the normalized embedding features and normalized weights respectively; No. Embedded features Hedi Class weight The angle between them is expressed as: (13) in, ; Discriminative additional angle margin classification loss function Expressed as: (14) in, is the scale parameter, represents the additional angular margin, Represents embedded features and corresponding category weights The angle between Represents the number of sample instances in this batch.

6. The method for detecting benthic organisms based on deep discriminant representation learning under a long-tail distribution according to claim 1, characterized in that: The distribution loss module of the uneven embedding feature includes: The embedding feature center construction submodule is used to transfer the distribution information of the embedded features of marine benthic organisms from the implicit head class to the tail class, propose an uneven embedding feature distribution loss, enhance the diversity of the embedded features of the tail class of marine benthic organisms, and construct the embedded features of each category separately; Intra-class embedding feature distribution variance calculation submodule: used to calculate the embedding feature distribution variance of each category of embedding features constructed by the embedding feature center construction submodule; Uneven embedding feature distribution loss submodule: used for uneven embedding feature distribution loss function of all categories based on the embedding feature distribution variance calculated by the intra-class embedding feature distribution variance calculation submodule.

7. The method for detecting benthic organisms based on deep discriminant representation learning under a long-tail distribution according to claim 6, characterized in that: By migrating the distribution information of the embedded features of marine benthic organisms from the implicit head class to the tail class, an uneven embedded feature distribution loss is proposed to enhance the diversity of the embedded features of the tail class of marine benthic organisms. The process of constructing the embedded features of each category is as follows: Construct the embedding features of each category and calculate the corresponding embedding feature center, which is expressed as: (15) in, , 、 、 and Representing sea urchin, scallop, starfish and sea cucumber categories respectively. Representative Training After batches, the corresponding category Embedded features, represents the center of the embedded feature, Representative Training The number of samples of the corresponding category after batches; Adopt sliding average mechanism to make full use of Embedding feature centers of batches To update Embedding feature centers of batches , expressed as: (16) in, , Represents the class center update rate.

8. The method for detecting benthic organisms based on deep discriminant representation learning under a long-tail distribution according to claim 6, characterized in that: The process of constructing the embedding features of each category and calculating the embedding feature distribution variance of the embedding feature center construction submodule is as follows: The angle between the embedded feature and its corresponding category center is expressed as: (17) in, It's training After the batch Embedded features Its corresponding center The angle between Average values of angular distribution for different categories Expressed as: (18) Subsequently, the intra-class embedding feature distribution variance is expressed as: (19) in, 、 、 and Representative Training Variance of intra-class embedding feature distributions for sea urchins, scallops, starfish, and sea cucumbers after batches.

9. The method for detecting benthic organisms based on deep discriminant representation learning under a long-tail distribution according to claim 6, characterized in that: Based on the embedding feature distribution variance calculated by the intra-class embedding feature distribution variance calculation submodule, the process of the uneven embedding feature distribution loss function for all categories is as follows: The uneven embedding feature distribution loss is expressed as: (20) in, 、 、 and Representatives from the set The first category of embedded feature distribution variance, the second category of embedded feature distribution variance, the third category of embedded feature distribution variance and the fourth category of embedded feature distribution variance, 、 、 and is the corresponding number of instances, 、 、 and Represents the uneven embedding feature distribution loss of different categories; Uneven embedding feature distribution loss for all categories Expressed as: (21)。

Images