Marine benthos progressive detection enhancement method guided by channel-space fuzzy logic attention under detail loss

Abstract

The invention provides a channel-space fuzzy logic attention-guided marine benthos progressive detection enhancement method under detail loss. The method comprises the following steps: constructing a data set based on pictures of one or more marine benthos of sea urchins, scallops, starfishes and sea cucumbers; dividing the data set into a training set and a test set; constructing an ocean benthos progressive detection enhancement model including a detail loss lower channel and spatial fuzzy logic attention guidance; training the channel-space fuzzy logic attention guided marine benthos progressive detection enhancement model based on the training set data to obtain a trained channel-space fuzzy logic attention guided marine benthos progressive detection enhancement model; and inputting test set data into the trained channel-space fuzzy logic attention-guided marine benthos progressive detection enhancement model to realize detection and identification of sea urchin, scallop, starfish and sea cucumber marine benthos.

Description

Technical Field

[0001] The invention belongs to the technical field of underwater visual intelligent perception, and relates to a progressive detection enhancement method for marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss. Background Art

[0002] Compared with the terrestrial environment, the real underwater environment contains a large amount of dissolved matter and particulate matter. [1]-[3] , such as underwater micro-phytoplankton, soluble organic matter and non-algae particles, which directly change the original propagation direction of light and produce scattering phenomenon [4] , which causes the loss of detailed information of marine benthic organisms, and undoubtedly increases the difficulty of detecting fuzzy marine benthic organisms. Therefore, it is urgent to propose a high-precision detection method for fuzzy marine benthic organisms.

[0003] To address the above challenges, existing methods can be summarized as follows: integrating contextual features [5] , explore the attention mechanism [6] , build a multi-layer spatial pyramid [7] , jointly optimize image enhancement and object detection tasks [8] , fuzzy fusion feature map [9]-

[11] Specifically, in order to improve the detection accuracy of fuzzy marine benthic organisms, the literature [5] A multi-scale context feature fusion module and a multi-scale fuzzy sampling module were proposed to effectively construct the multi-scale features of marine benthic organisms.

[12] First, channel attention is used to enhance the feature maps related to marine benthic organisms and suppress the feature maps related to the seabed background; then, spatial attention is used to enhance the features of the area containing marine benthic organisms and suppress the features of the seabed background area; finally, the features of marine benthic organisms are enhanced from both the channel and spatial levels, improving the detection performance of blurred marine benthic organisms.

[13] A feature enhancement gating module is proposed to selectively suppress or enhance multi-layer features, reducing the impact of complex underwater environment on subsequent feature fusion; a feature dynamic fusion module is designed to aggregate adjacent layer features, and the relationship between the target scale and feature fusion in the input image is established; a spatial pyramid hybrid pooling module is developed to enhance the ability of deep convolutional neural networks to extract texture features and contour features, effectively reducing the missed detection rate of ambiguous marine benthic organisms.

[14] A sample-weighted hyperparameter network model based on void convolution was constructed to form multiple semantic feature maps, which enhanced the detection performance of fuzzy marine benthic organisms; the progressive learning model from clean samples to fuzzy noise samples was explored to improve the detection performance of fuzzy marine benthic organisms.

[0004] Generally speaking, jointly optimizing multiple different tasks can improve the generalization performance of convolutional neural network models.[8] A method for jointly optimizing image enhancement and marine benthic organism detection tasks is proposed, and a guidance model based on channel and position dual attention is constructed; the degradation process is inverted using an underwater imaging model to obtain potential clear images, thereby enhancing the ability to express the details of blurred marine benthic organisms.

[15] A deep convolutional neural network for image enhancement-guided marine benthic organism detection is proposed. The core idea is to enhance the branch guidance to optimize the detection network, so that the detection branch features have both good visual effects and robust expression capabilities. The results show that this method can achieve superior detection results for fuzzy marine benthic organisms. In the weakly supervised learning mode, the literature

[16] A novel fuzzy marine benthic organism detection method is proposed. The core idea is to train two deep learning detectors simultaneously and design a batch sample selection strategy to encourage mutual learning between detectors, thereby improving the detection performance of fuzzy marine benthic organisms.

[0005] Through the review and analysis of existing work, the following deficiencies still exist in the fuzzy marine benthic organism detection:

[0006] (1) Using only a single discriminative channel descriptor or a texture channel descriptor cannot comprehensively enhance the informative feature map, and a single hidden layer feedforward neural network is difficult to specifically and efficiently learn the aforementioned discriminative channel descriptor and texture channel descriptor.

[0007] (2) Based on the channel cascade feature map, it is difficult to simultaneously enhance the regional responses of blurred marine benthic organisms at different scales using a single-scale convolution kernel.

[0008] (3) When a large amount of fuzzy perceptual information is introduced into the underwater optical imaging system, it is often difficult to derive appropriate channel attention weights or spatial attention weights by directly fusing the discriminative and texture channel attention maps or spatial attention maps with different receptive fields in an algebraic sum manner. Summary of the invention

[0009] In order to solve the above problems, the technical solution adopted by the present invention is: a method for progressive detection and enhancement of marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss, characterized in that it includes the following steps:

[0010] Construct a dataset based on images of one or more of the following marine benthic organisms: sea urchins, scallops, starfish, and sea cucumbers;

[0011] Divide the data set into: training set and test set;

[0012] A progressive detection enhancement model for marine benthic organisms is constructed, which includes a detail loss channel and spatial fuzzy logic attention guidance.

[0013] Based on the training set data, the channel-spatial fuzzy logic attention-guided marine benthic organisms progressive detection enhancement model is trained to obtain the trained channel-spatial fuzzy logic attention-guided marine benthic organisms progressive detection enhancement model;

[0014] The test set data is input into the trained channel-spatial fuzzy logic attention guided progressive detection enhancement model of marine benthic organisms to achieve detection and recognition of marine benthic organisms such as sea urchins, scallops, starfish and sea cucumbers.

[0015] Further: the progressive detection enhancement model of marine benthic organisms guided by the detail loss channel and spatial fuzzy logic attention includes:

[0016] Backbone network: extracts three different fine-grained feature maps from images of marine benthic organisms;

[0017] Detection head network: Based on the three different fine-grained feature maps output by the backbone network, feature fusion operations are performed to achieve the detection of large-sized, medium-sized and small-sized marine benthic organisms.

[0018] Further: the backbone network includes a first CBR3 module, a second CBR3 module, a CSFLA-1 module, a third CBR3 module, a CSFLA-2 module, a fourth CBR3 module, a first CSFLA-8 module, a fifth CBR3 module, a second CSFLA-8 module, a sixth CBR3 module and a CSFLA-4 module;

[0019] The first CBR3 module, the second CBR3 module, the CSFLA-1 module, the third CBR3 module, the CSFLA-2 module, the fourth CBR3 module, the first CSFLA-8 module, the fifth CBR3 module, the second CSFLA-8 module, the sixth CBR3 module and the CSFLA-4 module are connected in sequence;

[0020] The first CSFLA-8 module output, the second CSFLA-8 module output and the CSFLA-4 module output three different fine-grained feature maps respectively.

[0021] Further: the CSFLA-1 module, CSFLA-2 module, first CSFLA-8 module, second CSFLA-8 module and CSFLA-4 module represent CSFLA modules reused 1, 2, 8, 8 and 4 times in sequence respectively;

[0022] The CSFLA module includes 3×3 convolution, 1×1 convolution, fuzzy channel attention FCA module and fuzzy spatial attention FSA module, and adopts a skip connection strategy to prevent semantic information degradation;

[0023] The fuzzy channel attention FCA module is used to combine the spatial global average pooling and maximum pooling methods to construct the discriminative and textural channel descriptors about the fuzzy marine benthic features, and the channel fuzzy inference system is designed to fuse the discriminative and textural channel attention maps derived from two single hidden layer networks;

[0024] The fuzzy spatial attention FSA module constructs discriminative and textural spatial descriptors about fuzzy marine benthic features with the help of channel global average pooling and maximum pooling methods, and designs a spatial fuzzy reasoning system to fuse spatial attention maps with different receptive fields.

[0025] Further: the discriminative and texture channel descriptor construction process is as follows:

[0026] Two independent single hidden layer feedforward neural networks with the same network structure are used to learn the discriminative and texture channel descriptors derived from spatial global average pooling and maximum pooling operations respectively. The output of spatial global average pooling is expressed as:

[0027]

[0028] Among them, y a,k represents the average value of the feature map of the kth layer, i = 1, 2, ..., h, j = 1, 2, ..., w, h and w are the height and width of the current feature map respectively, represents the input features, represents the k-th layer feature map, represents the discriminative channel descriptor, k represents the feature layer index ranging from 1 to n c ;

[0029] The output of a single hidden layer feed-forward network that learns a discriminative channel descriptor is expressed as:

[0030] o a =s(W a,2 r(W a,1 y a )) (2)

[0031] in, represents the discriminative channel attention map, and Represent the hidden layer weights and output layer weights respectively, r(·) represents the Leaky ReLU activation function, and s(·) represents the activation function of o a Sigmoid activation function limited to the range (0,1);

[0032] Accordingly, the output of the spatial global maximum pooling operation (i.e., the texture channel descriptor) is expressed as:

[0033] y m，k=max{X k (1, 1), …, X k (h, w)} (3)

[0034] in, represents the texture channel descriptor, max{·} represents the maximum value operation, y m,k Represents the maximum value in the feature map of the kth layer;

[0035] Similarly, the output of a single hidden layer feed-forward network for learning a texture channel descriptor is expressed as:

[0036] o m =s(W m,2 r(W m,1 y m )) (4)

[0037] in, represents the texture channel attention map, and Represent the hidden layer and output layer weights respectively, and the activation function s(·) converts o m Clamped to the range (0,1).

[0038] Further: the design process of the channel fuzzy inference system is as follows:

[0039] Fusion of discriminative channel attention maps o a and texture channel attention map o m , the input variables and output variables are fuzzified by three fuzzy sets, namely "S", "M" and "L" (i.e. small, medium and large), and are characterized by Gaussian membership functions, expressed as:

[0040]

[0041] Among them, Θ∈{a,m,c}, μ Θ,k,S (·),μ Θ,k,M (·) and μ Θ,k,L (·) represents the small, medium and large membership of the channel attention map, σ represents the standard deviation, exp(·) represents the natural exponential function, c S = 0 represents the Gaussian membership function center of the fuzzy set as "small", c M = 0.5 represents the center of the Gaussian membership function of the fuzzy set as “medium”, c L =1.0 represents the Gaussian membership function center of the fuzzy set as “large”;

[0042] The following fuzzy rules are established:

[0043] Rule 1: If o a,k It is S and o m,k is S, then oc,k is S;

[0044] Rule 2: If o a,k It is S and o m,k is M, then o c,k is S;

[0045] Rule 3: If o a,k It is S and o m,k is L, then o c,k is M;

[0046] Rule 4: If o a,k It is M and o m,k is S, then o c,k is S;

[0047] Rule 5: If o a,k It is M and o m,k is M, then o c,k is M;

[0048] Rule 6: If o a,k It is M and o m,k is L, then o c,k is M;

[0049] Rule 7: If o a,k It is L and o m,k is S, then o c,k is M;

[0050] Rule 8: If o a,k It is L and o m,k is M, then o c,k is M;

[0051] Rule 9: If o a,k It is L and o m,k is L, then o c,k is L;

[0052] Among them, a,k represents the weight of the kth layer in the discriminative channel attention map, o m,k represents the weight of the kth layer in the texture channel attention map, o c,k Represents the k-th layer weight in the channel attention map after fuzzy reasoning;

[0053] Fuzzy reasoning is performed using the smaller operation, and the aggregate credibility (i.e., the union of the credibility reasoning results of each rule) is obtained by taking the larger operation, which is expressed as:

[0054] μ agg，k (o c，k )=max{min{μ a，k，Γ (o a，k ), μm，k，Γ (o m，k ), μ c，k，Γ (o c，k )}|n cfr =1, 2, ..., 9}(8)

[0055] Among them, Γ∈{S,M,L}, n cfr =1,…,9 represents the fuzzy rule index;

[0056] Deblurring is performed using the centroid method, expressed as:

[0057]

[0058] in, Represents the recalibration weight of the kth channel.

[0059] Furthermore, the channel features are recalibrated as follows:

[0060] The output of the FCA module is represented as:

[0061]

[0062] in, represents the recalibrated channel weights, B(·) represents broadcasting a one-dimensional vector into a three-dimensional matrix, and “⊙” represents element-by-element multiplication. Represents the features recalibrated using the FCA module.

[0063] Further: the discriminative and texture space descriptor construction process is as follows:

[0064] The channel global average pooling is expressed as:

[0065]

[0066] Where i = 1, ..., h, j = 1, ..., w, represents the discriminative spatial descriptor, n c Represents the number of channel layers;

[0067] The channel global maximum pooling is expressed as:

[0068]

[0069] in, stands for Texture Space Descriptor;

[0070] Then, the channel cascade feature map is expressed as:

[0071]

[0072] in, Represents the channel cascade feature map, symbol Represents the concatenation of features in the channel dimension;

[0073] Subsequently, the convolution operation under different receptive fields is expressed as:

[0074]

[0075] Among them, Ψ∈{p,q,l}, W p,i , W q,i and W l,i Represent the weights of the i-th layer in the convolution kernels of 3×3, 5×5, and 7×7, respectively. Spatial attention maps representing different receptive fields, Represents the convolution operation;

[0076] In order to enhance the nonlinear mapping capability, the spatial attention map is activated, which is expressed as:

[0077] O p =s(U p ) (15)

[0078] O q =s(U q ) (16)

[0079] O l =s(U l ) (17)

[0080] in, Represents the spatial attention map after activation.

[0081] Further: The design process of the spatial fuzzy reasoning system is as follows:

[0082] The activated spatial attention map O is fused using a fuzzy inference system. p , O q and O l , the input variables and output variables are fuzzified by three fuzzy sets, namely "S", "M" and "L" (i.e. small, medium and large), and described by Gaussian membership functions, expressed as:

[0083]

[0084] Among them, Φ∈{p,q,l,s}, μ Φ,i,j,S (·),μ Φ,i,j,M (·) and μ Φ,i,j,L (·) represents the small, medium, and large membership of the spatial attention map, respectively;

[0085] Create the following fuzzy rules:

[0086] Rule 1: If op,i,j It is S, o q,i,j It is S and o l,i,j is S, then o s,i,j is S;

[0087] Rule 2: If o p,i,j It is S, o q,i,j It is S and o l,i,j is M, then o s,i,j is S;

[0088] Rule 3: If o p,i,j It is S, o q,i,j It is S and o l,i,j is L, then o s,i,j is S;

[0089] Rule 4: If o p,i,j It is S, o q,i,j It is M and o l,i,j is S, then o s,i,j is S;

[0090] Rule 5: If o p,i,j It is S, o q,i,j It is M and o l,i,j is M, then o s,i,j is S;

[0091] Rule 6: If o p,i,j It is S, o q,i,j It is M and o l,i,j is L, then o s,i,j is M;

[0092] Rule 7: If o p,i,j It is S, o q,i,j It is L and o l,i,j is S, then o s,i,j is S;

[0093] Rule 8: If o p,i,j It is S, o q,i,j It is L and o l,i,j is M, then o s,i,j is M;

[0094] Rule 9: If o p,i,j It is S, o q,i,j It is L and o l,i,j is L, then o s,i,j is M;

[0095] Rule 10: If o p,i,j It is M, o q,i,j It is S and o l,i,j is S, then o s,i,j is S;

[0096] Rule 11: If o p,i,j It is M, o q,i,j It is S and o l,i,j is M, then o s,i,j is S;

[0097] Rule 12: If o p,i,j It is M, o q,i,j It is S and o l,i,j is L, then o s,i,j is M;

[0098] Rule 13: If o p,i,j It is M, o q,i,j It is M and o l,i,j is S, then o s,i,j is S;

[0099] Rule 14: If o p,i,j It is M, o q,i,j It is M and o l,i,j is M, then o s,i,j is M;

[0100] Rule 15: If o p,i,j It is M, o q,i,j It is M and o l,i,j is L, then o s,i,j is M;

[0101] Rule 16: If o p,i,j It is M, o q,i,j It is L and o l,i,j is S, then o s,i,j is M;

[0102] Rule 17: If o p,i,j It is M, o q,i,j It is L and o l,i,j is M, then o s,i,j is M;

[0103] Rule 18: If o p,i,j It is M, o q,i,j It is L and o l,i,j is L, then o s,i,j is L;

[0104] Rule 19: If o p,i,j Yes L, o q,i,j It is S and o l,i,j is S, then o s,i,j is S;

[0105] Rule 20: If o p,i,j Yes L, o q,i,jIt is S and o l,i,j is M, then o s,i,j is M;

[0106] Rule 21: If o p,i,j It is L, o q,i,j It is S and o l,i,j is L, then o s,i,j is M;

[0107] Rule 22: If o p,i,j It is L, o q,i,j It is M and o l,i,j is S, then o s,i,j is M;

[0108] Rule 23: If o p,i,j It is L, o q,i,j It is M and o l,i,j is M, then o s,i,j is M;

[0109] Rule 24: If o p,i,j It is L, o q,i,j It is M and o l,i,j is L, then o s,i,j is L;

[0110] Rule 25: If o p,i,j It is L, o q,i,j It is L and o l,i,j is S, then o s,i,j is M;

[0111] Rule 26: If o p,i,j It is L, o q,i,j It is L and o l,i,j is M, then o s,i,j is L;

[0112] Rule 27: If o p,i,j It is L, o q,i,j It is L and o l,i,j is L, then o s,i,j is L;

[0113] Similarly, fuzzy reasoning is performed by taking the smaller operation, and then the aggregate credibility (i.e., the union of the credibility reasoning results of each rule) in the spatial fuzzy reasoning system is obtained by taking the larger operation, which is expressed as:

[0114]

[0115] Among them, Γ∈{S,M,L}, n sfr =1,…,27 represents the fuzzy rule index;

[0116] Deblurring is performed using the centroid method, expressed as:

[0117]

[0118] in, Represents the recalibration weight at coordinate position (i, j).

[0119] The output of the FSA module is represented as:

[0120]

[0121] in, represents the recalibrated spatial weight, B(·) represents the operation of broadcasting a two-dimensional matrix into a three-dimensional matrix, Represents the features recalibrated using the FSA module.

[0122] A progressive detection enhancement method for marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss: including:

[0123] Building Module I: Building a dataset based on images of one or more marine benthic organisms including sea urchins, scallops, starfish, and sea cucumbers;

[0124] Partition module: divide the data set into training set and test set;

[0125] Building Module II: Constructing a channel-spatial fuzzy logic attention-guided progressive detection enhancement model for marine benthic organisms;

[0126] Training module: Based on the training set data, the channel-space fuzzy logic attention-guided marine benthic organism progressive detection enhancement model is trained to obtain the trained channel-space fuzzy logic attention-guided marine benthic organism progressive detection enhancement model;

[0127] Detection module: The test set data is input into the trained channel-spatial fuzzy logic attention-guided progressive detection enhancement model of marine benthic organisms to realize the detection and recognition of marine benthic organisms such as sea urchins, scallops, starfish and sea cucumbers.

[0128] The present invention provides a method for progressive detection enhancement of marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss, which has the following advantages:

[0129] 1. Combining spatial global average pooling and maximum pooling techniques, a discriminative and textural channel descriptor for fuzzy marine benthic features was constructed. With the help of a fuzzy inference system, the discriminative and textural channel attention maps were fused, a fuzzy channel attention module was proposed, and a comprehensive channel weight was derived. The feature map response related to fuzzy marine benthic organisms was enhanced, the seabed background noise feature map was suppressed, and the detection performance of fuzzy marine benthic organisms was improved from the channel level.

[0130] 2. A discriminative and texture spatial descriptor was constructed by combining channel global average pooling and maximum pooling techniques. The spatial attention maps with different receptive fields were fused with the help of the fuzzy inference system, and a fuzzy spatial attention module was proposed. Comprehensive spatial weights were derived, which enhanced the regional response of fuzzy marine benthic organisms and improved the detection performance of fuzzy marine benthic organisms from the spatial level.

[0131] 3. Combining the fuzzy channel attention module and the fuzzy spatial attention module, a fuzzy-channel controlled attention method with a residual mechanism is proposed, which suppresses the seabed background noise from both the channel and spatial levels, while enhancing the characteristics of marine benthic organisms, ultimately improving the detection performance of marine benthic organisms. BRIEF DESCRIPTION OF THE DRAWINGS

[0132] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. 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 creative labor.

[0133] Figure 1 A progressive detection enhancement model for marine benthic organisms guided by channel-spatial fuzzy logic attention;

[0134] Figure 2 is the fuzzy channel attention module;

[0135] Figure 3 is the fuzzy spatial attention module;

[0136] Figure 4 The performance comparison diagram of the accuracy-recall curves of different categories, where (a) is the accuracy-recall curve of sea urchins; (b) is the accuracy-recall curve of scallops, (c) is the accuracy-recall curve of starfish, and (d) is the accuracy-recall curve of sea cucumbers.

[0137] Figure 5 The following are qualitative comparisons of detection performance in typical scenarios, where (a) is a dark light scene, (b) is a bright light scene, (c) is a multi-scale target coexistence scene, and (d) is a blurred scene. DETAILED DESCRIPTION

[0138] It should be noted that, in the absence of 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.

[0139] 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 by no means 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 creative work are within the scope of protection of the present invention.

[0140] A method for progressive detection enhancement of marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss, comprising the following steps:

[0141] S1: Construct a dataset based on images of one or more marine benthic organisms including sea urchins, scallops, starfish, and sea cucumbers;

[0142] S2: Divide the data set into training set and test set;

[0143] S3: Construct a channel-spatial fuzzy logic attention-guided progressive detection enhancement model for marine benthic organisms;

[0144] S4: Based on the training set data, the channel-space fuzzy logic attention-guided marine benthic organism progressive detection enhancement model is trained to obtain the trained channel-space fuzzy logic attention-guided marine benthic organism progressive detection enhancement model;

[0145] S5: Input the test set data into the trained channel-spatial fuzzy logic attention-guided progressive detection enhancement model of marine benthic organisms to realize the detection and recognition of marine benthic organisms such as sea urchins, scallops, starfish and sea cucumbers.

[0146] The steps S1 / S2 / S3 / S4 / S5 are performed sequentially;

[0147] Figure 1 A progressive detection enhancement model for marine benthic organisms guided by channel-spatial fuzzy logic attention;

[0148] The channel-spatial fuzzy logic attention-guided marine benthic organism progressive detection enhancement model includes:

[0149] Backbone network: extracts three different fine-grained feature maps of 13×13×27, 26×26×27 and 52×52×27 for images of marine benthic organisms (resolution 416×416×3);

[0150] Detection head network: Based on the three different fine-grained feature maps output by the backbone network, feature fusion operations are performed to achieve the detection of large-sized, medium-sized and small-sized marine benthic organisms.

[0151] The backbone network includes a first CBR3 module, a second CBR3 module, a CSFLA-1 module, a third CBR3 module, a CSFLA-2 module, a fourth CBR3 module, a first CSFLA-8 module, a fifth CBR3 module, a second CSFLA-8 module, a sixth CBR3 module and a CSFLA-4 module;

[0152] The first CBR3 module, the second CBR3 module, the CSFLA-1 module, the third CBR3 module, the CSFLA-2 module, the fourth CBR3 module, the first CSFLA-8 module, the fifth CBR3 module, the second CSFLA-8 module, the sixth CBR3 module and the CSFLA-4 module are connected in sequence;

[0153] The first CSFLA-8 module output, the second CSFLA-8 module output and the CSFLA-4 module output three different fine-grained feature maps respectively.

[0154] The first CBR3 module, the second CBR3 module, the third CBR3 module, the fourth CBR3 module, the fifth CBR3 module, and the sixth CBR3 module have the same structure; the CBR3 module is constructed by 3×3 convolution, batch normalization, and ReLU activation function;

[0155] Further: the CSFLA-1 module, CSFLA-2 module, first CSFLA-8 module, second CSFLA-8 module and CSFLA-4 module respectively represent CSFLA modules reused 1, 2, 8, 8 and 4 times in sequence;

[0156] The CSFLA module includes 3×3 convolution, 1×1 convolution, fuzzy channel attention FCA module and fuzzy spatial attention FSA module, and adopts a skip connection strategy to prevent semantic information degradation;

[0157] Using the residual learning paradigm, combined with the fuzzy channel attention module and fuzzy spatial attention module constructed above, a progressive detection enhancement method for marine benthic organisms guided by Channel-Spatial Fuzzy Logic Attention (CSFLA) was finally systematically proposed.

[0158] The fuzzy channel attention FCA (Fuzzy Channel Attention) module is used to combine the spatial global average pooling and maximum pooling methods to construct the discriminative and textural channel descriptors about the characteristics of fuzzy marine benthic organisms, and the channel fuzzy inference system is designed to fuse the discriminative and textural channel attention maps derived from two single hidden layer networks;

[0159] The fuzzy spatial attention (FSA) module constructs discriminative and textural spatial descriptors about fuzzy marine benthic features with the help of channel global average pooling and maximum pooling methods, and designs a spatial fuzzy reasoning system that integrates spatial attention maps with different receptive fields.

[0160] The construction process of the fuzzy channel attention FCA (Fuzzy Channel Attention) module is as follows:

[0161] The construction process of the discriminative and texture channel descriptors is as follows:

[0162] Two independent single-hidden-layer feedforward neural networks with the same network structure are used to learn the discriminative and texture channel descriptors derived from spatial global average pooling and maximum pooling operations, respectively. Figure 2 As shown, the output of spatial global average pooling is expressed as:

[0163]

[0164] Among them, y a,k represents the average value of the feature map of the kth layer, i = 1, 2, ..., h, j = 1, 2, ..., w, h and w are the height and width of the current feature map respectively, represents the input features, represents the k-th layer feature map, Stands for Discriminative Channel Descriptor.

[0165] The output of a single hidden layer feed-forward network that learns a discriminative channel descriptor is expressed as:

[0166] o a =s(W a,2 r(W a,1 y a )) (2)

[0167] in, represents the discriminative channel attention map, and Represent the hidden layer weights and output layer weights respectively, r(·) represents the Leaky ReLU activation function, and s(·) represents the activation function of oa Sigmoid activation function constrained to the range (0,1).

[0168] Accordingly, the output of the spatial global maximum pooling operation (i.e., the texture channel descriptor) is expressed as:

[0169] y m,k =max{X k (1,1),L,X k (h,w)} (3)

[0170] in, represents the texture channel descriptor, max{·} represents the maximum value operation, y m,k Represents the maximum value in the feature map of the kth layer.

[0171] Similarly, the output of a single hidden layer feed-forward network for learning a texture channel descriptor is expressed as:

[0172] o m =s(W m,2 r(W m,1 y m )) (4)

[0173] in, represents the texture channel attention map, and Represent the hidden layer and output layer weights respectively, and the activation function s(·) converts o m Clamped to the range (0,1).

[0174] The design process of the channel fuzzy inference system is as follows:

[0175] Inspired by the application of fuzzy inference systems in clustering, classification and segmentation tasks, we use fuzzy inference systems to fuse discriminative channel attention maps. a and texture channel attention map o m Specifically, the input variables and output variables are fuzzified by three fuzzy sets, namely "S", "M" and "L" (i.e. small, medium and large), and are characterized by Gaussian membership functions, expressed as:

[0176]

[0177] Among them, Θ∈{a,m,c}, μ Θ,k,S (·),μ Θ,k,M (·) and μ Θ,k,L (·) represents the small, medium and large membership of the channel attention map, σ represents the standard deviation, exp(·) represents the natural exponential function, c S =0, c M = 0.5 and c L=1.0 represents the corresponding fuzzy set center.

[0178] The following fuzzy rules are established:

[0179] Rule 1: If o a,k It is S and o m,k is S, then o c,k is S;

[0180] Rule 2: If o a,k It is S and o m,k is M, then o c,k is S;

[0181] Rule 3: If o a,k It is S and o m,k is L, then o c,k is M;

[0182] Rule 4: If o a,k It is M and o m,k is S, then o c,k is S;

[0183] Rule 5: If o a,k It is M and o m,k is M, then o c,k is M;

[0184] Rule 6: If o a,k It is M and o m,k is L, then o c,k is M;

[0185] Rule 7: If o a,k It is L and o m,k is S, then o c,k is M;

[0186] Rule 8: If o a,k It is L and o m,k is M, then o c,k is M;

[0187] Rule 9: If o a,k It is L and o m,k is L, then o c,k is L;

[0188] Fuzzy reasoning is performed using the smaller operation (i.e., the premises are operated with an “and” operation to obtain the conclusion), and the aggregate credibility is obtained by taking the larger operation (i.e., the union of the credibility reasoning results of each rule), which is expressed as:

[0189] μ agg，k (o c，k )=max{min{μ a，k，Γ (o a，k), μ m，k，Γ (o m，k ), μ c，k，Γ (o c，k )}|n cfr =1, 2, ..., 9}(8)

[0190] Among them, Γ∈{S,M,L}, n cfr =1,…,9 represents the fuzzy rule index.

[0191] Deblurring is performed using the centroid method, expressed as:

[0192]

[0193] in, Represents the recalibration weight of the kth channel.

[0194] Furthermore, the channel features are recalibrated as follows:

[0195] The output of the FCA module is represented as:

[0196]

[0197] in, represents the recalibrated channel weights, B(·) represents broadcasting a one-dimensional vector into a three-dimensional matrix, and “⊙” represents element-by-element multiplication. Represents the features recalibrated using the FCA module.

[0198] The design process of the fuzzy spatial attention module is as follows:

[0199] From a spatial perspective, the accuracy of fuzzy marine benthic organism detection depends on the response of the marine benthic organisms and the seabed background area in the feature map. In this case, the fuzzy spatial attention module is designed to adaptively adjust the spatial weights in the feature map. Figure 3 As shown, channel global average pooling and maximum pooling techniques are used to construct discriminative and texture spatial descriptors.

[0200] The construction process of discriminative and texture space descriptors is as follows:

[0201] The channel global average pooling is expressed as:

[0202]

[0203] Where i = 1, ..., h, j = 1, ..., w, stands for Discriminative Spatial Descriptor.

[0204] The channel global maximum pooling is expressed as:

[0205]

[0206] in, Stands for Texture Space Descriptor.

[0207] Then, the channel cascade feature map is expressed as:

[0208]

[0209] in, Represents the channel cascade feature map, symbol Represents concatenating features in the channel dimension.

[0210] Subsequently, the convolution operation under different receptive fields is expressed as:

[0211]

[0212] Among them, Ψ∈{p,q,l}, W p,i , W q,i and W l,i Represent the weights of the i-th layer in the convolution kernels of 3×3, 5×5, and 7×7, respectively. Spatial attention maps representing different receptive fields, Represents a convolution operation.

[0213] In order to enhance the nonlinear mapping capability, the spatial attention map is activated, which is expressed as:

[0214] O p =s(U p ) (15)

[0215] O q =s(U q ) (16)

[0216] O l =s(U l ) (17)

[0217] in, Represents the spatial attention map after activation.

[0218] The design process of spatial fuzzy reasoner is as follows:

[0219] The activated spatial attention map O is fused using a fuzzy inference system p , O q and O l Specifically, the input and output variables are fuzzified by three fuzzy sets, namely "S", "M" and "L" (i.e. small, medium and large), and described by Gaussian membership functions, expressed as:

[0220]

[0221] Among them, Φ∈{p,q,l,s}, μ Φ,i,j,S (·),μ Φ,i,j,M (·) and μ Φ,i,j,L (·) represent the small, medium, and large membership of the spatial attention map, respectively.

[0222] Create the following fuzzy rules:

[0223] Rule 1: If o p,i,j It is S, o q,i,j It is S and o l,i,j is S, then o s,i,j is S;

[0224] Rule 2: If o p,i,j It is S, o q,i,j It is S and o l,i,j is M, then o s,i,j is S;

[0225] Rule 3: If o p,i,j It is S, o q,i,j It is S and o l,i,j is L, then o s,i,j is S;

[0226] Rule 4: If o p,i,j It is S, o q,i,j It is M and o l,i,j is S, then o s,i,j is S;

[0227] Rule 5: If o p,i,j It is S, o q,i,j It is M and o l,i,j is M, then o s,i,j is S;

[0228] Rule 6: If o p,i,j It is S, o q,i,j It is M and o l,i,j is L, then o s,i,j is M;

[0229] Rule 7: If o p,i,j It is S, o q,i,j It is L and o l,i,j is S, then o s,i,j is S;

[0230] Rule 8: If o p,i,j It is S, o q,i,j It is L and o l,i,j is M, then o s,i,j is M;

[0231] Rule 9: If o p,i,jIt is S, o q,i,j It is L and o l,i,j is L, then o s,i,j is M;

[0232] Rule 10: If o p,i,j It is M, o q,i,j It is S and o l,i,j is S, then o s,i,j is S;

[0233] Rule 11: If o p,i,j It is M, o q,i,j It is S and o l,i,j is M, then o s,i,j is S;

[0234] Rule 12: If o p,i,j It is M, o q,i,j It is S and o l,i,j is L, then o s,i,j is M;

[0235] Rule 13: If o p,i,j It is M, o q,i,j It is M and o l,i,j is S, then o s,i,j is S;

[0236] Rule 14: If o p,i,j It is M, o q,i,j It is M and o l,i,j is M, then o s,i,j is M;

[0237] Rule 15: If o p,i,j It is M, o q,i,j It is M and o l,i,j is L, then o s,i,j is M;

[0238] Rule 16: If o p,i,j It is M, o q,i,j It is L and o l,i,j is S, then o s,i,j is M;

[0239] Rule 17: If o p,i,j It is M, o q,i,j It is L and o l,i,j is M, then o s,i,j is M;

[0240] Rule 18: If o p,i,j It is M, o q,i,j It is L and o l,i,j is L, then o s,i,j is L;

[0241] Rule 19: If o p,i,j It is L, o q,i,j It is S and o l,i,j is S, then o s,i,j is S;

[0242] Rule 20: If o p,i,j It is L, o q,i,j It is S and o l,i,j is M, then o s,i,j is M;

[0243] Rule 21: If o p,i,j It is L, o q,i,j It is S and o l,i,j is L, then o s,i,j is M;

[0244] Rule 22: If o p,i,j It is L, o q,i,j It is M and o l,i,j is S, then o s,i,j is M;

[0245] Rule 23: If o p,i,j It is L, o q,i,j It is M and o l,i,j is M, then o s,i,j is M;

[0246] Rule 24: If o p,i,j It is L, o q,i,j It is M and o l,i,j is L, then o s,i,j is L;

[0247] Rule 25: If o p,i,j It is L, o q,i,j It is L and o l,i,j is S, then o s,i,j is M;

[0248] Rule 26: If o p,i,j It is L, o q,i,j It is L and o l,i,j is M, then o s,i,j is L;

[0249] Rule 27: If o p,i,j It is L, o q,i,j It is L and o l,i,j is L, then o s,i,j is L;

[0250] Similarly, fuzzy reasoning is performed by taking the smaller operation (i.e., the “and” operation is performed between the premises to obtain the rule conclusion), and then the aggregate credibility in the spatial fuzzy reasoning system is obtained by taking the larger operation (i.e., the union of the credibility reasoning results of each rule), which is expressed as:

[0251]

[0252] Among them, Γ∈{S,M,L}, n sfr =1,…,27 represents the fuzzy rule index.

[0253] Deblurring is performed using the centroid method, expressed as:

[0254]

[0255] in, Represents the recalibration weight at coordinate position (i, j).

[0256] The spatial response recalibration is as follows:

[0257] The output of the FSA module is represented as:

[0258]

[0259] in, represents the recalibrated spatial weight, B(·) represents the operation of broadcasting a two-dimensional matrix into a three-dimensional matrix, Represents the features recalibrated using the FSA module.

[0260] The detection head network includes a first detection head sub-network, a second detection head sub-network, and a third detection head sub-network;

[0261] The third detection head sub-network includes a CBR5 module, a CBR3 module, and a CBR1 module;

[0262] The CBR1 module is constructed by convolution (1×1), batch normalization, and ReLU activation function;

[0263] The CBR3 module is constructed by convolution (3×3), batch normalization, and ReLU activation function;

[0264] The CBR5 module is composed of a CBR1 module, a CBR3 module, a CBR1 module, a CBR3 module and a CBR1 module in sequence;

[0265] The first detection head sub-network and the second detection head sub-network have the same structure;

[0266] The first detection head sub-network includes a first CU module and a first detection sub-network;

[0267] The output ends of the first CSFLA-8 module and the first CU module are connected to the first detection sub-network through splicing;

[0268] The first detection sub-network has the same structure as the third detection head sub-network;

[0269] The first detection head sub-network includes a second CU module and a second detection sub-network;

[0270] The output ends of the second CSFLA-8 module and the second CU module are connected to the second detection sub-network through splicing;

[0271] The CBR5 module input end of the third detection head sub-network is connected to the input end of the CSFLA-4 module, and the CBR5 module output end is connected to the input end of the second CU module and the CBR3 module input end of the third detection head sub-network;

[0272] The CU module is constructed by convolution, batch normalization, ReLU activation function and upsampling, and its function is to increase the resolution of feature maps. In order to enhance the generalization of the detection model, two different fine-grained feature maps are spliced ​​in the channel dimension;

[0273] For an input image with a resolution of 416×416×3, feature maps with dimensions of 13×13×27, 26×26×27, and 52×52×27 can be obtained, which are used to detect large-, medium-, and small-scale marine benthic organisms, respectively.

[0274] Embodiment 1:

[0275] 1. Dataset and experimental configuration;

[0276] The present invention verifies the effectiveness and superiority of the CSFLA method on the URPC2019 and DUO datasets, both of which include four categories (i.e., sea urchins, scallops, starfish, and sea cucumbers). The URPC2019 and DUO datasets contain 4757 and 7782 images, respectively. Regarding the URPC2019 dataset, the training set contains 3806 images and the test set contains 951 images; regarding 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 present invention uses AP, mAP@.5, mAP@.75, and mAP@[.5,.95] as detection performance evaluation indicators. Using a five-fold cross-validation strategy, the average of the five test results is taken as the detection accuracy of different models; a video with a resolution of 1280×720 is used to obtain the inference speed of the model.

[0277] Table 1

[0278]

[0279] 2. Selection of fuzzy membership width parameter value

[0280] In order to determine the optimal value of the fuzzy membership function width parameter σ, the present invention selects from the candidate set {0.05, 0.1, 0.2, 0.3, 0.4, 0.5}. It can be seen from Tables 2 and 3 that σ = 0.05 or σ = 0.1 makes the designed membership function too narrow, resulting in poor detection performance, while σ = 0.4 or σ = 0.5 makes the designed membership function too wide, which also leads to poor detection performance. The results show that it is more desirable to choose σ = 0.2 and σ = 0.3. When σ = 0.2, the most superior mAP@.75 and mAP@[.5,.95] indicator values ​​can be achieved on both the URPC2019 dataset and the DUO dataset. For convenience, the present invention finally sets the parameter σ to 0.2.

[0281] Table 2

[0282]

[0283] Table 3

[0284]

[0285] 3. Fuzzy attention performance

[0286] In order to demonstrate the effectiveness and superiority of the FCA and FSA modules in fusing different types of channel attention maps and different receptive field spatial attention maps, the present invention considers the algebraic sum-based channel attention (ASCA) method and the algebraic sum-based spatial attention (ASSA) method. For ease of comparison, the present invention uses the YOLOv3 model to obtain the benchmark detection accuracy.

[0287] Tables 4 and 5 summarize the detection results on the URPC2019 dataset and the DUO dataset, respectively. Regarding the detection performance on the URPC2019 dataset, it can be seen from Table 4 that the detection accuracy of the baseline model is significantly lower than that of the detection methods based on ASCA and ASSA. Specifically, in terms of mAP@.5, mAP@.75, and mAP@[.5,.95] indicators, the detection accuracy based on ASCA is 1.17%, 1.29%, and 0.67% higher than that of the baseline model, respectively, indicating that the ASCA module can enhance the characteristics of blurred marine benthic organisms from the channel level while suppressing the seabed background noise. Similarly, the detection accuracy based on the ASSA module is 0.91%, 1.20%, and 1.65% higher than that of the baseline model, respectively, indicating that the ASSA module can enhance the regional response of blurred marine benthic organisms from the spatial level. More importantly, the proposed FCA and FSA modules achieve higher detection accuracy than the ASCA and ASSA modules in terms of mAP@.5, mAP@.75, and mAP@[.5,.95] index values, which indicates that with the help of the fuzzy inference system, the FCA and FSA modules can more robustly enhance the fuzzy marine benthic features and suppress the seabed background noise from the channel and spatial levels, respectively. Regarding the DUO dataset, it can be seen from Table 5 that except for the mAP@.75 and sea urchin index values, the proposed FCA and FSA modules achieve more superior detection performance than the ASCA and ASSA modules.

[0288] Table 4

[0289]

[0290] Table 5

[0291]

[0292] 4. Quantitative comparison of classic attention detection methods

[0293] In order to prove the superiority of the CSFLA method, the present invention conducts comprehensive comparative experiments with SENet, CBAM and SKNet methods on the URPC2019 and DUO datasets. Tables 6 and 7 summarize the corresponding comparative results. It can be seen that the detection accuracy of the SENet model is significantly lower than that of the proposed CSFLA method. The main reason is that the SENet model only uses the spatial global average pooling technology to extract discriminative information, and it is difficult to construct a comprehensive channel descriptor. In addition, although the CBAM model uses both spatial attention and channel attention techniques, its detection performance is still weaker than that of the proposed CSFLA method. The important reasons are summarized as follows: the number of hidden layer nodes in the channel attention is less than that of the input layer and the output layer, which destroys the direct connection between the channel weight and the corresponding feature map; single-scale convolution kernels are difficult to simultaneously generate large-scale, medium-scale and small-scale regional responses. In addition, although SKNet can adaptively select the receptive field of the convolution kernel, its detection accuracy is still lower than that of the proposed CSFLA method. The main reason is that the SKNet model adopts algebraic sum strategy to fuse feature maps with different receptive fields. Especially when the visual imaging system introduces fuzzy perceptual information, the robustness and generalization of the SKNet model are poor.

[0294] Table 6

[0295]

[0296] Table 7

[0297]

[0298] Figure 4 The performance comparison diagram of the accuracy-recall curves of different categories, where (a) is the accuracy-recall curve of sea urchins; (b) is the accuracy-recall curve of scallops, (c) is the accuracy-recall curve of starfish, and (d) is the accuracy-recall curve of sea cucumbers.

[0299] In addition, the present invention carried out a precision-recall performance comparison experiment. Figure 4It can be seen that by combining channel attention and spatial attention in turn, the CBAM model has better detection performance than SENet. The main reason is that SENet only enhances the feature map of marine benthic organisms at the channel level, while the CBAM model enhances the feature map and regional response related to marine benthic organisms at the channel and spatial levels respectively. In addition, the detection performance of the SKNet model exceeds that of the CBAM model. The key reason is that the SKNet model can adaptively select the receptive field and realize multi-scale feature extraction. Overall, the proposed CSFLA method can achieve the optimal balance between accuracy and recall. The main reasons include: 1) The FCA module recalibrates the channel weights and improves the detection performance of fuzzy marine benthic organisms from the channel level; 2) The FSA module recalibrates the spatial weights, suppresses the seabed background noise, and improves the detection performance of fuzzy marine benthic organisms from the spatial level.

[0300] 5. Quantitative comparison of special marine benthic organism detection methods

[0301] In order to further verify the superiority of the CSFLA method, the present invention comprehensively considers the dedicated marine benthic detection methods including TVFRD, Pruning-SSD, Quantized-SSD, YOLOv3-UW, DG-YOLO and BBROABR. Tables 8 and 9 summarize the experimental results of different algorithms on the URPC2019 and DUO datasets, respectively. It can be seen from Table 8 that at a resolution of 608×608, the proposed CSFLA method can not only achieve the most superior mAP@.5 and mAP@[.5,.95] index values, but also achieve the best performance when detecting sea urchins. With the help of generalized intersection-over-union and marine benthic prior anchor boxes, the BBROABR method can comprehensively consider the positioning performance and recall performance, so the BBROABR method can achieve the best performance when detecting scallops and starfish. However, the BBROABR method lacks the attention mechanism to enhance the fuzzy marine benthic features, and it is difficult to obtain the highest comprehensive evaluation index value (i.e., mAP@[.5,.95]). In addition, by using the domain invariant module, DG-YOLO can achieve the best performance in detecting sea cucumbers. Due to the lack of spatial attention to enhance the spatial response of marine benthic organisms, the YOLOv3-UW method using only channel attention is difficult to achieve the best detection performance. Using the depth-separable convolution strategy and weight mapping technology, the Quantized-SSD method achieved the fastest detection speed (31FPS) when the input image resolution was 512×512. By using the convolution kernel pruning technology to remove low-contribution neurons, the Pruning-SSD method without channel attention and spatial attention also found it difficult to achieve superior detection performance. The TVFRD method uses the region recommendation network to generate a large number of candidate detection boxes in advance. Although it improves the detection accuracy, it can only achieve a detection speed of 3FPS, which is difficult to meet the needs of real-time detection. Regarding the detection performance on the DUO dataset, it can be seen from Table 9 that, except for the sea cucumber category, the CSFLA method has achieved excellent detection performance in other indicators.

[0302] Table 8

[0303]

[0304]

[0305] Table 9

[0306]

[0307] 6. Quantitative comparison of common target detection methods

[0308] To illustrate the superiority of the CSFLA method, the present invention comprehensively considers two-stage detection methods (Faster R-CNN), single-stage detection methods (SSD, EfficientDet, YOLOv2, YOLOv3, YOLOv4, CEFP2N and M2YOLOF), and anchor-free detection methods (CenterNet and FCOS). Tables 10 and 11 summarize the detection performance of different methods on the URPC2019 and DUO datasets. It can be seen that the anchor-free CenterNet and FCOS methods are difficult to achieve superior detection performance. The main reason is that the prior anchor box information is not used during the training process. Although the use of context enhancement modules and feature purification modules can enhance texture information, the detection performance of the CEFP2N method is significantly weaker than that of the proposed CSFLA method. The key reason is that the CEFP2N method uses algebraic sum strategies to construct channel attention weights and spatial attention weights. In particular, when the visual imaging system introduces fuzzy perception information, it is difficult to derive appropriate channel weights and spatial weights. In addition, by designing a multi-input single-output encoder and a dynamic sample selection strategy, the M2YOLOF method effectively improves the global representation ability and enhances the quality of positive samples, thereby improving the detection performance of small targets on the general target detection dataset. Except for the mAP@.75 indicator, the detection performance of the M2YOLOF method is significantly weaker than that of the proposed CSFLA method, which shows that fuzzy channel attention and fuzzy spatial attention play an important role in enhancing the characteristics of marine benthic organisms and suppressing the seabed background noise. On the URPC2019 and DUO datasets, the YOLOv4 model can achieve the best detection accuracy for the sea urchin category. The main reason is that the YOLOv4 model proposes a Mosaic technology of randomly mixing 4 pictures, which is more conducive to improving the diversity of training samples, especially for categories with a large number of them. In addition, since the YOLOv3 model does not adopt the channel and spatial attention strategies, the feature maps related to fuzzy marine benthic organisms cannot be enhanced, so it is difficult to achieve superior detection performance. Compared with the use of the VGG-16 backbone network, SSD can achieve the fastest detection speed (i.e., 38FPS) using MobileNet v2. Different from the single-stage detection method, since Faster R-CNN needs to use the region recommendation network to generate a large number of candidate boxes that are not related to the category, the real-time performance is difficult to guarantee. Overall, the proposed CSFLA method achieves the best value in most indicators. The key reason is that the FCA and FSA modules derive more comprehensive channel and spatial weights.

[0309] Table 10

[0310]

[0311]

[0312] Table 11

[0313]

[0314] 7. Qualitative comparison of marine benthic organism detection

[0315] In order to verify the detection superiority of the proposed CSFLA method in real environments, this paper considers TVFRD, Pruning-SSD, Quantized-SSD, YOLOv3-UW and DG-YOLO methods, and the corresponding experiments are carried out in dark light scenes, bright light scenes, multi-scale target coexistence scenes and blurred scenes. Figure 5 The following are qualitative comparisons of detection performance in typical scenarios, where (a) is a dark light scene, (b) is a bright light scene, (c) is a multi-scale target coexistence scene, and (d) is a blurred scene.

[0316] like Figure 5 As shown in the figure, in low-light scenes, although the CSFLA method cannot detect all marine benthic organisms, its detection performance far exceeds that of other methods. The TVFRD method can achieve better detection performance in bright light scenes, mainly because the region recommendation network generates a large number of candidate regions in the first stage. Regarding the performance comparison in multi-scale target coexistence scenes, especially the detection of marine benthic organisms under seaweed occlusion, the proposed CSFLA method achieves the best detection performance, mainly because the CSFLA method suppresses the seabed background noise and enhances the characteristics of marine benthic organisms. Thanks to the generalization of the CSFLA method, almost consistent conclusions can be drawn in blurry scenes.

[0317] 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 aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned 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.

[0318] [1]Li C, Anwar S, Hou J, et al. Underwater image enhancement via mediumtransmission-guided multi-color space embedding[J]. IEEE TransactionsonImage Processing, 2021, 30: 4985-5000.

[0319] [2]Zhuang P,Wu J,Porikli F,et al.Underwater image enhancementwithhyper-laplacian reflectance priors[J].IEEE Transactions on ImageProcessing,2022,31:5442-5455.

[0321] [3]Wang N,Chen T,Kong X,et al.Underwater attentional generativeadversarialnetworks for image enhancement[J].IEEE Transactions on Human-MachineSystems,2023,53(3):490-500.

[0322] [4]Zhou J,Yang T,Chu W,et al.Underwater image restoration viabackscatter pixelprior and color compensation[J].Engineering Applications ofArtificialIntelligence,2022,111:104785.

[0323] [5]Liu C,Wang Z,Wang S,et al.A new dataset,Poisson GAN and AquaNetforunderwater object grabbing[J].IEEE Transactions on Circuits and SystemsforVideo Technology,2022,32(5):2831-2844.

[0324] [6]Chen W,Fan B.Underwater object detection with mixed attentionmechanismand multi-enhancement strategy[C].Chinese Automation Congress,Shanghai,China,2020:2821-2826.

[0326] [7]Peng F,Miao Z,Li F,et al.S-FPN:A shortcut feature pyramid networkfor seacucumber detection in underwater images[J].Expert Systems withApplications,2021,182:115306.

[0328] [8]Cheng N,Xie H,Zhu X,et al.Joint image enhancement learning formarineobject detection in natural scene[J].Engineering Applications ofArtificialIntelligence,2023,120:105905.

[0329] [9]Saha S,Bovolo F,Bruzzone L.Building change detection in VHR SARimagesvia unsupervised deep transcoding[J].IEEE Transactions on GeoscienceandRemote Sensing,2021,59(3):1917-1929.

[0330]

[10] Huang K,Zhang Y,Cheng H,et al.Semantic segmentation of breastultrasoundimage with fuzzy deep learning network and breast anatomyconstraints[J].

[0331] Neurocomputing,2021,450:319-335.

[0332]

[11] Zhou Y,Mao A,Huo S,et al.Salient object detection via fuzzytheory andobject-level enhancement[J].IEEE Transactions on Multimedia,2018,21(1):74-85.

[0334]

[12] Zhang J,Zhu L,Xu L,et al.MFFSSD:An enhanced SSD for underwaterobjectdetection[C].ChineseAutomation Congress,Shanghai,China,2020:5938-5943.

[13] Hua X,Cui X,Xu X,et al.Underwater object detection algorithm basedonfeature enhancement and progressive dynamic aggregation strategy[J].PatternRecognition,2023,139:109511.

[0335]

[14] Chen L,Zhou F,Wang S,et al.SWIPENET:Object detection in noisyunderwaterscenes[J].Pattern Recognition,2022,132:108926.

[0336]

[15] Liu Z,Zhuang Y,Jia P,et al.A novel underwater image enhancementalgorithmand an improved underwater biological detection pipeline[J].Journalof MarineScience and Engineering,2022,10(9):1204.

[0337]

[16] Cai S,Li G,Shan Y.Underwater object detection using collaborativeweaklysupervision[J].Computers and Electrical Engineering,2022,102:108159。

Claims

1. A method for progressive detection enhancement of marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss, characterized by: The following steps are involved: Construct a dataset based on images of one or more of the following marine benthic organisms: sea urchins, scallops, starfish, and sea cucumbers; Divide the data set into: training set and test set; A progressive detection enhancement model for marine benthic organisms is constructed, which includes a detail loss channel and spatial fuzzy logic attention guidance. Based on the training set data, the channel-spatial fuzzy logic attention-guided marine benthic organisms progressive detection enhancement model is trained to obtain the trained channel-spatial fuzzy logic attention-guided marine benthic organisms progressive detection enhancement model; The test set data is input into the trained channel-spatial fuzzy logic attention guided progressive detection enhancement model of marine benthic organisms to achieve detection and recognition of marine benthic organisms such as sea urchins, scallops, starfish and sea cucumbers.

2. The method for progressive detection enhancement of marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss according to claim 1, characterized in that: The progressive detection enhancement model of marine benthic organisms guided by the channel under detail loss and spatial fuzzy logic attention includes: Backbone network: extracts three different fine-grained feature maps from images of marine benthic organisms; Detection head network: Based on the three different fine-grained feature maps output by the backbone network, feature fusion operations are performed to achieve the detection of large-sized, medium-sized and small-sized marine benthic organisms.

3. The method for progressive detection enhancement of marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss according to claim 1, characterized in that: The backbone network includes a first CBR3 module, a second CBR3 module, a CSFLA-1 module, a third CBR3 module, a CSFLA-2 module, a fourth CBR3 module, a first CSFLA-8 module, a fifth CBR3 module, a second CSFLA-8 module, a sixth CBR3 module and a CSFLA-4 module; The first CBR3 module, the second CBR3 module, the CSFLA-1 module, the third CBR3 module, the CSFLA-2 module, the fourth CBR3 module, the first CSFLA-8 module, the fifth CBR3 module, the second CSFLA-8 module, the sixth CBR3 module and the CSFLA-4 module are connected in sequence; The first CSFLA-8 module output, the second CSFLA-8 module output and the CSFLA-4 module output three different fine-grained feature maps respectively.

4. The method for progressive detection enhancement of marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss according to claim 3, characterized in that: The CSFLA-1 module, CSFLA-2 module, first CSFLA-8 module, second CSFLA-8 module and CSFLA-4 module represent CSFLA modules that are reused 1, 2, 8, 8 and 4 times in sequence, respectively; The CSFLA module includes 3×3 convolution, 1×1 convolution, fuzzy channel attention FCA module and fuzzy spatial attention FSA module, and adopts a skip connection strategy to prevent semantic information degradation; The fuzzy channel attention FCA module is used to combine the spatial global average pooling and maximum pooling methods to construct the discriminative and textural channel descriptors about the fuzzy marine benthic features, and the channel fuzzy inference system is designed to fuse the discriminative and textural channel attention maps derived from two single hidden layer networks; The fuzzy spatial attention FSA module constructs discriminative and textural spatial descriptors about fuzzy marine benthic features with the help of channel global average pooling and maximum pooling methods, and designs a spatial fuzzy reasoning system to fuse spatial attention maps with different receptive fields.

5. The method for progressive detection enhancement of marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss according to claim 3, characterized in that: The discriminative and texture channel descriptors are constructed as follows: Two independent single hidden layer feedforward neural networks with the same network structure are used to learn the discriminative and texture channel descriptors derived from spatial global average pooling and maximum pooling operations respectively. The output of spatial global average pooling is expressed as: Among them, y a,k represents the average value of the feature map of the kth layer, i = 1, 2, ..., h, j = 1, 2, ..., w, h and w are the height and width of the current feature map respectively, represents the input features, represents the k-th layer feature map, represents the discriminative channel descriptor, k represents the feature layer index ranging from 1 to n c ; The output of a single hidden layer feed-forward network that learns a discriminative channel descriptor is expressed as: o a =s(W a,2 r(W a,1 y a )) (2) in, represents the discriminative channel attention map, and Represent the hidden layer weights and output layer weights respectively, r(·) represents the Leaky ReLU activation function, and s(·) represents the activation function of o a Sigmoid activation function limited to the range (0,1); Accordingly, the output of the spatial global maximum pooling operation, the texture channel descriptor, is expressed as: y m，k＝ max{X k (1,1),…,X k (h,w)} (3) in, represents the texture channel descriptor, max{·} represents the maximum value operation, y m,k Represents the maximum value in the feature map of the kth layer; Similarly, the output of a single hidden layer feed-forward network for learning a texture channel descriptor is expressed as: o m =s(W m,2 r(W m,1 y m )) (4) in, represents the texture channel attention map, and Represent the hidden layer and output layer weights respectively, and the activation function s(·) converts o m Clamped to the range (0,1).

6. The method for progressive detection enhancement of marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss according to claim 3, characterized in that: The design process of the channel fuzzy inference system is as follows: Fusion of discriminative channel attention maps o a and texture channel attention map o m , the input variables and output variables are fuzzified by three fuzzy sets, namely "S", "M" and "L", i.e. small, medium and large, and are characterized by Gaussian membership functions, expressed as: Among them, Θ∈{a,m,c}, μ Θ,k,S (·),μ Θ,k,M (·) and μ Θ,k,L (·) represents the small, medium and large membership of the channel attention map, σ represents the standard deviation, exp(·) represents the natural exponential function, c S =0 represents the Gaussian membership function center of the fuzzy set as "small", c M = 0.5 represents the center of the Gaussian membership function of the fuzzy set as "medium", c L =1.0 represents the Gaussian membership function center of the fuzzy set as "large"; The following fuzzy rules are established: Rule 1: If o a,k It is S and o m,k is S, then o c,k is S; Rule 2: If o a,k It is S and o m,k is M, then o c,k is S; Rule 3: If o a,k It is S and o m,k is L, then o c,k is M; Rule 4: If o a，k It is M and o m，k is S, then o c，k is S; Rule 5: If o a，k It is M and o m，k is M, then o c，k is M; Rule 6: If o a，k It is M and o m，k is L, then o c，k is M; Rule 7: If o a，k It is L and o m，k is S, then o c，k is M; Rule 8: If o a，k It is L and o m，k is M, then o c，k is M; Rule 9: If o a，k It is L and o m，k is L, then o c，k is L; Among them, a，k represents the weight of the kth layer in the discriminative channel attention map, o m，k represents the weight of the kth layer in the texture channel attention map, o c，k Represents the k-th layer weight in the channel attention map after fuzzy reasoning; Fuzzy reasoning is performed using the smaller operation, and the aggregated credibility is obtained by taking the larger operation, which is the union of the credibility reasoning results of each rule, expressed as: m agg，k (the c，k )=max{min{μ a，kΓ (the a，k ),m m，k，Γ (the m，k ),m c，k，Γ (the c，k )}|n cfr =1,2,…,9} (8) Among them, Γ∈{S,M,L},n cfr =1,…,9 represents the fuzzy rule index; Deblurring is performed using the centroid method, expressed as: in, Represents the recalibration weight of the kth channel. Furthermore, the channel features are recalibrated as follows: The output of the FCA module is represented as: in, represents the recalibrated channel weight, B(·) represents broadcasting a one-dimensional vector into a three-dimensional matrix, "⊙" represents element-by-element multiplication, Represents the features recalibrated using the FCA module.

7. The method for progressive detection and enhancement of marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss according to claim 1, characterized in that: The discriminative and texture space descriptors are constructed as follows: The channel global average pooling is expressed as: Where i = 1, ..., h, j = 1, ..., w, represents the discriminative spatial descriptor, n c Represents the number of channel layers; The channel global maximum pooling is expressed as: in, stands for Texture Space Descriptor; Then, the channel cascade feature map is expressed as: in, Represents the channel cascade feature map, symbol Represents the concatenation of features in the channel dimension; Subsequently, the convolution operation under different receptive fields is expressed as: Among them, Ψ∈{p,q,l}, W p,i , W q,i and W l,i Represent the weights of the i-th layer in the convolution kernels of 3×3, 5×5, and 7×7, respectively. Spatial attention maps representing different receptive fields, Represents the convolution operation; In order to enhance the nonlinear mapping capability, the spatial attention map is activated, which is expressed as: THE p =s(U p ) (15) THE q =s(U q ) (16) THE l =s(U l ) (17) in, Represents the spatial attention map after activation.

8. The method for progressive detection enhancement of marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss according to claim 1, characterized in that: The design process of the spatial fuzzy reasoning system is as follows: The activated spatial attention map O is fused using a fuzzy inference system. p , O q and O l , the input variables and output variables are fuzzified by three fuzzy sets, namely "S", "M" and "L" (i.e. small, medium and large), and described by Gaussian membership functions, expressed as: Among them, Φ∈{p,q,l,s}, μ Φ,i,j,S (·),μ Φ,i,j,M (·) and μ Φ,i,j,L (·) represents the small, medium, and large membership of the spatial attention map, respectively; Create the following fuzzy rules: Rule 1: If o p,i,j It is S, o q,i,j It is S and o l,i,j is S, then o s,i,j is S; Rule 2: If o p,i,j It is S, o q,i,j It is S and o l,i,j is M, then o s,i,j is S; Rule 3: If o p,i,j It is S, o q,i,j It is S and o l,i,j is L, then o s,i,j is S; Rule 4: If o p,i,j It is S, o q,i,j It is M and o l,i,j is S, then o s,i,j is S; Rule 5: If o p,i,j It is S, o q,i,j It is M and o l,i,j is M, then o s,i,j is S; Rule 6: If o p,i,j It is S, o q,i,j It is M and o l,i,j is L, then o s,i,j is M; Rule 7: If o p,i,j It is S, o q,i,j It is L and o l,i,j is S, then o s,i,j is S; Rule 8: If o p,i,j It is S, o q,i,j It is L and o l,i,j is M, then o s,i,j is M; Rule 9: If o p,i,j It is S, o q,i,j It is L and o l,i,j is L, then o s,i,j is M; Rule 10: If o p,i,j It is M, o q,i,j It is S and o l,i,j is S, then o s,i,j is S; Rule 11: If o p,i,j It is M, o q,i,j It is S and o l,i,j is M, then o s,i,j is S; Rule 12: If o p,i,j It is M, o q,i,j It is S and o l,i,j is L, then o s,i,j is M; Rule 13: If o p,i,j It is M, o q,i,j It is M and o l,i,j is S, then o s,i,j is S; Rule 14: If o p,i,j It is M, o q,i,j It is M and o l,i,j is M, then o s,i,j is M; Rule 15: If o p,i,j It is M, o q,i,j It is M and o l,i,j is L, then o s,i,j is M; Rule 16: If o p,i,j It is M, o q,i,j It is L and o l,i,j is S, then o s,i,j is M; Rule 17: If o p,i,j It is M, o q,i,j It is L and o l,i,j is M, then o s,i,j is M; Rule 18: If o p,i,j It is M, o q,i,j It is L and o l,i,j is L, then o s,i,j is L; Rule 19: If o p,i,j It is L, o q,i,j It is S and o l,i,j is S, then o s,i,j is S; Rule 20: If o p,i,j It is L, o q,i,j It is S and o l,i,j is M, then o s,i,j is M; Rule 21: If o p,i,j It is L, o q,i,j It is S and o l,i,j is L, then o s,i,j is M; Rule 22: If o p,i,j It is L, o q,i,j It is M and o l,i,j is S, then o s,i,j is M; Rule 23: If o p,i,j It is L, o q,i,j It is M and o l,i,j is M, then o s,i,j is M; Rule 24: If o p,i,j It is L, o q,i,j It is M and o l,i,j is L, then o s,i,j is L; Rule 25: If o p,i,j It is L, o q,i,j It is L and o l,i,j is S, then o s,i,j is M; Rule 26: If o p,i,j It is L, o q,i,j It is L and o l,i,j is M, then o s,i,j is L; Rule 27: If o p,i,j It is L, o q,i,j It is L and o l,i,j is L, then o s,i,j is L; Similarly, fuzzy reasoning is performed by taking the smaller operation, and then the aggregate credibility in the spatial fuzzy reasoning system is obtained by taking the larger operation, which is the union of the credibility reasoning results of each rule, expressed as: Among them, Γ∈{S,M,L}, n sfr =1,…,27 represents the fuzzy rule index; Deblurring is performed using the centroid method, expressed as: in, Represents the recalibration weight at coordinate position (i, j). The output of the FSA module is represented as: in, represents the recalibrated spatial weight, B(·) represents the operation of broadcasting a two-dimensional matrix into a three-dimensional matrix, Represents the features recalibrated using the FSA module.

9. A method and device for progressive detection enhancement of marine benthic organisms guided by channel-spatial fuzzy logic attention under detail loss, characterized by: include: Building Module I: Building a dataset based on images of one or more marine benthic organisms including sea urchins, scallops, starfish, and sea cucumbers; Partition module: divide the data set into training set and test set; Building Module II: Constructing a channel-spatial fuzzy logic attention-guided progressive detection enhancement model for marine benthic organisms; Training module: Based on the training set data, the channel-spatial fuzzy logic attention-guided marine benthic organisms progressive detection enhancement model is trained to obtain the trained channel-spatial fuzzy logic attention-guided marine benthic organisms progressive detection enhancement model; Detection module: The test set data is input into the trained channel-spatial fuzzy logic attention-guided progressive detection enhancement model of marine benthic organisms to realize the detection and recognition of marine benthic organisms such as sea urchins, scallops, starfish and sea cucumbers.

Images