A method and system for detecting shrimp fry in aquaculture without anchor

Abstract

The application discloses an aquaculture shrimp fry anchor-free detection method and system, and relates to the field of deep learning target detection. The method comprises the following steps: obtaining picture data sets of to-be-detected shrimp fries collected on site at an aquaculture base and inputting the picture data sets into a feature extraction backbone network of a coordinate attention mechanism module to obtain a plurality of coordinate attention processed feature maps; inputting the plurality of coordinate attention processed feature maps into a neck feature fusion network of a self-adaptive spatial feature fusion module to obtain a plurality of self-adaptive fusion feature maps; and inputting the fusion feature maps into a head regression detection network adopting a generalized intersection over union module to obtain an anchor-free detection result of the shrimp fries. In the anchor-free detection process of the shrimp fries, the coordinate attention module, the self-adaptive spatial feature fusion module and the generalized intersection over union module are introduced in the feature extraction stage, the feature fusion stage and the regression detection stage, respectively, so that accurate detection of the aquaculture shrimp fries can be realized.

Description

Technical Field

[0001] This invention relates to the field of deep learning object detection, and in particular to a method and system for detecting anchorless shrimp larvae in intensive aquaculture. Background Technology

[0002] As people's living standards improve, the demand for seafood is also increasing. According to statistics from the "State of Fisheries and Aquaculture in the World" published by the Food and Agriculture Organization of the United Nations, as early as 2020, the total output of fisheries and aquaculture exceeded 200 million tons, with the largest catch being the whiteleg shrimp. As an important component of aquatic products, shrimp not only provide humans with high-quality nutritional resources but also play a vital role in global food security. Pollution, overfishing, and climate change have become significant factors restricting the development of marine fisheries, constantly threatening human food security. Aquaculture, on the other hand, can avoid the impact of the natural environment, providing humans with higher-quality and more stable shrimp products, and contributing to a virtuous cycle of the ecosystem.

[0003] Aquaculture is an important industry in my country, and intensive farming has become an inevitable trend in modern fisheries development. In intensive shrimp farming, shrimp larvae testing is an indispensable part, forming the basis for shrimp larvae counting, disease identification, and high-quality breeding. Therefore, accurate and efficient shrimp larvae testing plays a crucial role in shrimp behavior analysis, biomass estimation, and automated growth monitoring, providing vital support for intensive shrimp farming.

[0004] End-to-end detection methods based on deep learning can automatically extract target features from large amounts of images, reducing labor costs. Compared to traditional machine learning or manual methods, deep learning-based detection methods, with their stable and efficient advantages, are gradually replacing manual operations, providing a new paradigm for automated detection of shrimp in aquaculture.

[0005] However, current research on shrimp detection based on deep learning mainly focuses on adult shrimp, and methods used for detecting adult shrimp cannot be directly applied to detecting shrimp larvae. Compared to detecting adult shrimp, detecting shrimp larvae is more difficult. On the one hand, shrimp larvae are small in size, containing less feature information. In convolutional neural networks, as downsampling operations occur, target feature information is easily lost and contaminated by conflicting background information. On the other hand, due to the shrimp larvae's own activity patterns, they often move in a hopping manner, and coupled with the influence of the farming environment, captured images of shrimp larvae often appear blurry. Although deep learning-based shrimp detection has made some progress in adult shrimp, overcoming the interference caused by the shrimp larvae themselves and environmental factors, and improving the performance of shrimp larvae detection, remains a pressing problem to be solved in intensive aquaculture. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for detecting shrimp larvae without anchors in aquaculture. In the process of detecting shrimp larvae without anchors, a coordinate attention module, an adaptive spatial feature fusion module, and a generalized intersection-union-comparison module are introduced in the feature extraction stage (backbone network), the feature fusion stage (neck), and the regression detection stage (head), respectively, so as to achieve accurate detection of shrimp larvae in aquaculture.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] This invention provides a method for detecting anchorlessness in aquaculture shrimp larvae, the method comprising:

[0009] Obtain the dataset of shrimp larvae images to be detected;

[0010] The dataset of shrimp larvae images to be detected is input into an improved feature extraction backbone network to obtain multiple feature maps with coordinate attention processing. The improved feature extraction backbone network is a traditional feature extraction backbone network with a coordinate attention mechanism module connected to the output of each output branch.

[0011] Multiple feature maps processed by the coordinate attention method are input into an improved neck feature fusion network to obtain multiple adaptive fused feature maps. The improved neck feature fusion network employs an adaptive spatial feature fusion module within a traditional neck feature fusion network. The adaptive spatial feature fusion module and the adaptive fused feature map are then integrated. Figure 1 One-to-one correspondence;

[0012] The adaptive fusion feature map is input into the improved head regression detection network to obtain the shrimp larvae anchorless detection result; the improved head regression detection network is to replace the traditional cross-union module of the traditional head regression detection network with the generalized cross-union module.

[0013] Optionally, the dataset of shrimp larvae images to be detected is input into an improved feature extraction backbone network, specifically including:

[0014] The dataset of shrimp larvae images to be detected is input into the improved feature extraction backbone network to obtain the feature maps to be processed generated by the layer before each coordinate attention mechanism module;

[0015] For each feature map to be processed, each channel is encoded along the horizontal and vertical directions of the feature map to obtain horizontal encoded features and vertical encoded features.

[0016] The horizontal and vertical coding features are concatenated to obtain an intermediate feature map containing both horizontal and vertical coding information.

[0017] The intermediate feature map is split into two independent tensors to obtain a horizontal tensor and a vertical tensor;

[0018] The horizontal tensor and the vertical tensor are respectively converted into tensors with the same number of channels as the feature map to be processed, to obtain the horizontal target tensor and the vertical target tensor;

[0019] The horizontal target tensor and the vertical target tensor are applied to each position of each channel of the feature map to be processed to obtain the output feature map processed by the coordinate attention mechanism module; the output feature map is the feature map processed by the coordinate attention mechanism.

[0020] Optionally, the expression for the horizontal directional encoding feature is:

[0021]

[0022] In the formula, This represents the encoded feature obtained at height h of the c-th channel in the horizontal direction; W represents the width of the feature map to be processed. This represents the i-th feature value in the h-th row of the c-th channel of the feature map to be processed;

[0023] The expression for the vertical direction encoded feature is:

[0024]

[0025] In the formula, This represents the encoded feature obtained at the width w of the c-th channel in the vertical direction; H represents the height of the feature map to be processed. This represents the j-th feature value in the w-th column of the c-th channel of the feature map to be processed;

[0026] The expression for the intermediate feature map is:

[0027]

[0028] Where f is an intermediate feature map with horizontal and vertical encoded information, and z h This indicates that the feature map to be processed has been generated after all channels of the feature map have been encoded with features in the horizontal direction; z w This indicates that the feature map to be processed is generated after all channels of the feature map have been encoded with features in the vertical direction; [z] h , z w ] indicates that z h and z w Perform concatenation operations in spatial dimensions; F1 is a transformation function implemented using 1×1 convolution; It is a non-linear activation function;

[0029] The expression for the horizontal target tensor is:

[0030]

[0031] Among them, g h Represents the horizontal target tensor; F represents the tensor containing horizontally encoded information obtained from the splitting; h This represents a transformation function implemented using a 1×1 convolution; Represents the Sigmoid function;

[0032] The expression for the vertical target tensor is:

[0033]

[0034] Among them, g w Represents the vertical target tensor; F represents the tensor containing vertically encoded information obtained from the splitting; w This represents a transformation function implemented using a 1×1 convolution;

[0035] The expression for the output feature map is:

[0036]

[0037] Among them, y c (i, j) represents the new feature value at position (i, j) of the c-th channel of the feature map to be processed after being processed by the coordinate attention mechanism; x c (i, j) represents the feature value at position (i, j) of the c-th channel of the feature map to be processed; The coordinate attention weight represents the position at the i-th location within the height h of the c-th channel. The attention weight represents the coordinates at the j-th position within the c-th channel width w.

[0038] Optionally, the feature maps from the coordinate attention processing are input into an improved neck feature fusion network to obtain multiple adaptive fusion feature maps, specifically including:

[0039] The feature maps of the coordinate attention processing are input into the improved neck feature fusion network to obtain the fusion feature map to be processed generated by the previous layer of each adaptive spatial feature fusion module; each fusion feature map to be processed is a fusion feature map at a different level; the different levels of the fusion feature map to be processed are divided according to different resolutions;

[0040] The fusion feature maps to be processed at each level are input into each of the adaptive spatial feature fusion modules, and the fusion weights at each position of the fusion feature maps to be processed at each level are adaptively calculated; and the predicted feature map of each level is obtained based on the fusion weights and the fusion feature maps to be processed at each level; the predicted feature map is the adaptive fusion feature map.

[0041] Optionally, the adaptive fusion expression of the adaptive spatial feature fusion module is:

[0042]

[0043] In the formula, This represents the adaptive fusion feature at position (i, j) in the nth layer predicted feature map; ; This represents the feature vector at position (i, j) in the fused feature map of layer m to be processed, which is then adjusted to the fused feature map of layer n to be processed. These are the fusion weights when the three different levels of the fusion feature map to be processed are adjusted to the nth level of the fusion feature map to be processed; n=1, 2, 3; m=1, 2, 3.

[0044] Optionally, the calculation function expression of the generalized intersection-union ratio module is:

[0045]

[0046] in,

[0047] In the formula, GIoU represents the generalized intersection-union ratio; A p Represents the predicted bounding box; A g A represents the true bounding box of the target; c Representing A p and A g The minimum bounding rectangle;

[0048] The loss function expression for the generalized intersection-union module is:

[0049] .

[0050] Optionally, the improved head regression detection network employs a decoupled prediction head structure with three parallel branches;

[0051] The decoupled prediction head structure includes a classification branch for target classification, a regression branch for predicting bounding box location coordinates, and a confidence branch for calculating target confidence.

[0052] This invention also provides a system for detecting anchorless shrimp larvae in aquaculture, the system comprising:

[0053] The image acquisition unit is used to acquire images of shrimp larvae to be tested, collected on-site from aquaculture bases.

[0054] The feature extraction unit is used to input the shrimp seedling image to be detected into the improved feature extraction backbone network to obtain multiple feature maps with coordinate attention processing; the improved feature extraction backbone network is a coordinate attention mechanism module connected to the output end of each output branch of the traditional feature extraction backbone network.

[0055] An adaptive feature fusion unit is used to input multiple feature maps processed by the coordinate attention method into an improved neck feature fusion network to obtain multiple adaptive fused feature maps. The improved neck feature fusion network employs an adaptive spatial feature fusion module within a traditional neck feature fusion network. The adaptive spatial feature fusion module and the adaptive fused feature map... Figure 1 One-to-one correspondence;

[0056] The shrimp larvae anchorless detection unit is used to input the adaptive fusion feature map into the improved head regression detection network to obtain the shrimp larvae anchorless detection result; the improved head regression detection network replaces the traditional cross-over and union ratio module of the traditional head regression detection network with the generalized cross-over and union ratio module.

[0057] Optionally, the feature extraction unit specifically includes:

[0058] The feature map acquisition subunit is used to input the shrimp seedling image to be detected into the improved feature extraction backbone network to obtain the feature map to be processed generated by the previous layer of each coordinate attention mechanism module;

[0059] The feature encoding subunit is used to encode each channel along the horizontal and vertical directions of each feature map to be processed, respectively, to obtain horizontally encoded features and vertically encoded features.

[0060] The splicing subunit is used to perform splicing operations based on the horizontal direction encoding features and the vertical direction encoding features to obtain an intermediate feature map with horizontal and vertical direction encoding information.

[0061] The splitting subunit is used to split the intermediate feature map into two independent tensors, resulting in a horizontal tensor and a vertical tensor.

[0062] The transformation subunit is used to convert the horizontal tensor and the vertical tensor into tensors with the same number of channels as the feature map to be processed, respectively, to obtain the horizontal target tensor and the vertical target tensor;

[0063] The coordinate attention application subunit is used to apply the horizontal target tensor and the vertical target tensor to each position of each channel of the feature map to be processed, so as to obtain the output feature map processed by the coordinate attention mechanism module; the output feature map is the feature map processed by the coordinate attention.

[0064] Optionally, the adaptive feature fusion unit specifically includes:

[0065] The subunit for obtaining the fused feature map to be processed is used to input the feature maps of multiple coordinate attention processing into the improved neck feature fusion network to obtain the fused feature map to be processed generated by the previous layer of each adaptive spatial feature fusion module; each of the fused feature maps to be processed is a fused feature map at a different level; the different levels of the fused feature map to be processed are divided according to different resolutions;

[0066] An adaptive fusion subunit is used to input the fusion feature maps to be processed at each level into each of the adaptive spatial feature fusion modules, adaptively calculate the fusion weight at each position of the fusion feature maps to be processed at each level, and obtain the predicted feature map at each level based on the fusion weight and the fusion feature maps to be processed at each level; the predicted feature map is the adaptive fusion feature map.

[0067] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0068] This invention proposes an anchorless detection method and system for aquaculture shrimp larvae. In the anchorless detection process, a coordinate attention module, an adaptive spatial feature fusion module, and a generalized intersection-union-comparison (IUCN) module are introduced in the feature extraction stage (backbone network), the feature fusion stage (neck), and the regression detection stage (head). This enhances the feature extraction capability for fuzzy, small-sized shrimp larvae targets, reduces the interference of conflicting background information on detection, and better optimizes bounding box regression, ultimately achieving accurate detection of aquaculture shrimp larvae. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0070] Figure 1 This is a flowchart of a method for detecting anchorless shrimp larvae in aquaculture, provided in Embodiment 1 of the present invention.

[0071] Figure 2This is a diagram showing the overall structure of the target detection model for anchorless detection of shrimp larvae in aquaculture, as provided in Embodiment 1 of the present invention.

[0072] Figure 3 This is a schematic diagram of the components of the target detection model for anchorless detection of shrimp larvae in aquaculture provided in Embodiment 1 of the present invention.

[0073] Figure 4 This is a schematic diagram of the coordinate attention mechanism module structure provided in Embodiment 1 of the present invention;

[0074] Figure 5 This is a schematic diagram of the adaptive spatial feature fusion module structure provided in Embodiment 1 of the present invention;

[0075] Figure 6 This is a schematic diagram comparing the traditional cross-union ratio and the generalized cross-union ratio provided in Embodiment 1 of the present invention;

[0076] Figure 7 This is a schematic diagram of the decoupling detection head structure provided in Embodiment 1 of the present invention. Detailed Implementation

[0077] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0078] The purpose of this invention is to provide a method and system for anchorless detection of shrimp larvae in aquaculture. By introducing a coordinate attention module, an adaptive spatial feature fusion module, and a generalized intersection-union-compare module into the feature extraction stage (backbone network), the feature fusion stage (neck), and the regression detection stage (head) of the shrimp larvae during the anchorless detection process, the method enhances its feature extraction capability for fuzzy, small-sized shrimp larvae targets, reduces the interference of conflicting information such as background on detection, and better optimizes bounding box regression, ultimately achieving accurate detection of shrimp larvae in aquaculture.

[0079] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0080] Example 1

[0081] like Figure 1 As shown in the figure, this embodiment provides a method for detecting anchorlessness in aquaculture shrimp larvae, the method comprising:

[0082] S1: Obtain the dataset of shrimp larvae images collected on-site from aquaculture bases.

[0083] In step S1, an appropriate amount of shrimp larvae were placed in a rearing tray to create an environment suitable for their survival. A camera was used to capture video footage of the larvae's normal activities from directly above. The target species were shrimp larvae approximately 20 days after hatching, randomly selected from the breeding workshop of an aquaculture base and placed in the rearing tray for video recording. Throughout the experiment, the water temperature was controlled at 23 degrees Celsius and the pH at approximately 8 to ensure no harm was caused to the larvae. For the collected shrimp larvae dataset, a Python program was used to split the video frame by frame into images, constructing a shrimp larvae dataset containing 680 images. All images in the dataset contain shrimp larvae, with an average of 50 images per image.

[0084] To ensure the accuracy of target detection, the target detection model used for the shrimp larvae images to be tested is generally a pre-trained target detection model. In order to facilitate the training of the target detection model, the LabelImg tool is used to label the shrimp larvae targets in the sample images, and the images are saved in VOC2007 format. The training set, validation set and test set are randomly divided in a ratio of 7:2:1.

[0085] S2: Input the shrimp seedling image dataset to be detected into the improved feature extraction backbone network to obtain multiple feature maps with coordinate attention processing; the improved feature extraction backbone network is a coordinate attention mechanism module connected to the output end of each output branch of the traditional feature extraction backbone network.

[0086] In this embodiment, Figure 2 This is the overall architecture diagram of the object detection model, which includes an improved feature extraction backbone network, an improved neck feature fusion network, and an improved head regression detection network. Figure 3 (a) is a structural diagram of the Concat and CBS components in the object detection model; Figure 3 (b) is a structural diagram of the CSPn component in the target detection model; Figure 3 (c) is a structural diagram of the Focus component in the object detection model; Figure 3 (d) is a structural diagram of the SPP component in the target detection model.

[0087] The feature extraction process is illustrated using the backbone network CSPDarknet as an example. The structural diagram of the backbone network CSPDarknet during the feature extraction stage is shown below. Figure 2As shown, there are three output branches. The first output branch, which includes the Focus layer, CBS layer (stride=2), CSP1 layer, CBS layer (stride=2), CSP3 layer, and CA layer, consists of the Focus layer, CBS layer (stride=2), CSP3 layer, and CA layer. The second output branch, which includes the CBS layer (stride=2), CSP3 layer, and CA layer, consists of the CBS layer (stride=2), SPP layer, CSP1 layer, and CA layer. The third output branch, which includes the CBS layer (stride=2), SPP layer, CSP1 layer, and CA layer, consists of the CBS layer (stride=2), SPP layer, CSP1 layer, and CA layer.

[0088] Among them, such as Figure 4 As shown, step S2 specifically includes:

[0089] S21: Input the dataset of shrimp seedling images to be detected into the improved feature extraction backbone network to obtain the feature maps to be processed generated by the layer before each coordinate attention mechanism module.

[0090] Corresponding feature map to be processed Figure 2 The feature map generated by the last CSP layer in each output branch. Figure 4 It can be seen that the feature map to be processed needs to undergo residual operation before feature encoding.

[0091] S22: For each feature map to be processed, encode each channel along the horizontal and vertical directions of the feature map to obtain horizontal coded features and vertical coded features.

[0092] Two 1D global pooling operations are used to encode each channel along the horizontal and vertical directions of the feature map, respectively, using pooling kernels of (H, 1) and (1, W). Figure 4 The X Avg Pool and Y Avg Pool operations are described below. "X Avg Pool" and "Y Avg Pool" refer to the one-dimensional horizontal global pool and the one-dimensional vertical global pool, respectively.

[0093] The expression for the horizontal direction encoding feature is as follows:

[0094] (1)

[0095] In the formula, This represents the encoded feature obtained at height h of the c-th channel in the horizontal direction; W represents the width of the feature map to be processed. This represents the i-th feature value in the h-th row of the c-th channel of the feature map to be processed;

[0096] The expression for the vertical direction encoded feature is:

[0097] (2)

[0098] In the formula, This represents the encoded feature obtained at the width w of the c-th channel in the vertical direction; H represents the height of the feature map to be processed. This represents the j-th feature value in the c-th channel w column of the feature map to be processed.

[0099] Through the two transformations in formulas (1) and (2) above, the attention block is allowed to capture remote dependencies in one spatial direction and retain precise location information in another direction, which helps the network to more accurately locate shrimp seedling objects.

[0100] S23: Perform a concatenation operation based on the horizontal and vertical coding features to obtain an intermediate feature map with horizontal and vertical coding information.

[0101] For z with global receptive field and precise location information h and z w Perform coordinate attention generation, i.e., the calculation process of formulas (3) to (5).

[0102] The expression for the intermediate feature map is:

[0103] (3)

[0104] In the formula, the generated It is an intermediate feature map containing horizontal and vertical encoded information, z h This indicates that the feature map to be processed has been generated after all channels of the feature map have been encoded with features in the horizontal direction; z w This indicates that the feature map to be processed is generated after all channels of the feature map have been encoded with features in the vertical direction; [z] h , z w ] indicates that z h and z w Perform concatenation operations in spatial dimensions; F1 is a transformation function implemented using 1×1 convolution; is a non-linear activation function. r is a hyperparameter, and we take r=32.

[0105] The specific operation corresponding to this step Figure 4 The Concat+Conv_2D and BatchNorm+Non-liner operations are described in the text. "Concat+Conv_2D" refers to the concatenation operation in spatial dimensions and the 1×1 convolution operation; "BatchNorm+Non-liner" refers to the batch normalization operation and the non-linear activation operation.

[0106] S24: Split the intermediate feature map into two independent tensors to obtain a horizontal tensor and a vertical tensor.

[0107] The intermediate feature map f is split into two independent tensors, namely the horizontal tensor. and vertical tensor .correspond Figure 4 The split operation in [the context of the text].

[0108] S25: Convert the horizontal tensor and the vertical tensor into tensors with the same number of channels as the feature map to be processed, respectively, to obtain the horizontal target tensor and the vertical target tensor.

[0109] The expression for the horizontal target tensor is:

[0110] (4)

[0111] Among them, g h Represents the horizontal target tensor; F represents the tensor containing horizontally encoded information obtained from the splitting; h This represents a transformation function implemented using a 1×1 convolution; Represents the Sigmoid function;

[0112] The expression for the vertical target tensor is:

[0113] (5)

[0114] Among them, g w Represents the vertical target tensor; F represents the tensor containing vertically encoded information obtained from the splitting; w This represents a transformation function implemented using a 1×1 convolution;

[0115] Combination Figure 4 It can be seen that the split horizontal tensor and the split vertical tensor are respectively subjected to Conv_2D operation and Sigmoid operation.

[0116] S26: Apply the horizontal target tensor and the vertical target tensor to each position of each channel of the feature map to be processed, to obtain the output feature map processed by the coordinate attention mechanism module; the output feature map is the feature map processed by the coordinate attention mechanism. This step corresponds to... Figure 4 The re-weighting operation in [the context of the text].

[0117] Finally, the generated coordinate attention is applied to each position (determined by h and w) of each channel (denoted as c) of the feature map to be processed, that is, the expression of the output feature map is:

[0118] (6)

[0119] Among them, y c (i, j) represents the new feature value at position (i, j) of the c-th channel of the feature map to be processed after being processed by the coordinate attention mechanism; x c (i, j) represents the feature value at position (i, j) of the c-th channel of the feature map to be processed; The coordinate attention weight represents the position at the i-th location within the height h of the c-th channel. The attention weight represents the coordinates at the j-th position within the c-th channel width w.

[0120] In step S2, a coordinate attention mechanism module is added to the backbone network used for feature extraction to improve the feature representation of blurry, small-sized targets, thereby enhancing the method's feature extraction capability. CSPDarknet is used as the backbone network for target feature extraction. The backbone network uses operations such as convolution and pooling to output three feature maps with different resolutions: 80×80×128, 40×40×256, and 20×20×512, containing target location and semantic information. Figure 2 As shown, a coordinate attention mechanism is added before the three output branches to establish the dependencies between different positions on each channel of the feature map, so that the feature extraction method focuses on and extracts the features of the target region.

[0121] S3: Input the feature maps from the coordinate attention processing into the improved neck feature fusion network to obtain multiple adaptive fused feature maps; the improved neck feature fusion network is a traditional neck feature fusion network employing an adaptive spatial feature fusion module; the adaptive spatial feature fusion module and the adaptive fused feature map... Figure 1 One-to-one correspondence. Step S3 correspondence. Figure 2 The feature fusion stage in the process.

[0122] Specifically, step S3 includes:

[0123] S31: Input the feature maps of the coordinate attention processing into the improved neck feature fusion network to obtain the fusion feature map to be processed generated by the previous layer of each adaptive spatial feature fusion module; each fusion feature map to be processed is a fusion feature map at a different level; the different levels of the fusion feature map to be processed are divided according to different resolutions.

[0124] In step S2, the output branches of the three coordinate attention mechanism modules of the backbone network CSPDarknet each generate a coordinate attention-processed feature map. These three coordinate attention-processed feature maps obtained from feature extraction are then input into the improved neck feature fusion network. Figure 2 The CBS layer, UpSample layer, CSP1 layer, and shown in the figure The splicing operation represents the fusion structure and fusion process, resulting in three fusion feature maps to be processed, corresponding to... Figure 2 The feature maps output by the three grid shadow boxes are three different levels of feature maps to be fused, corresponding to... Figure 5 In the three feature layers X with different resolutions: level_1 Stride=32, level_2 Stride=16, and level_3 Stride=8,... 1 X 2 X 3 .

[0125] S32: Input the fusion feature maps to be processed at each level into each of the adaptive spatial feature fusion modules, adaptively calculate the fusion weight at each position of the fusion feature maps to be processed at each level; and obtain the predicted feature map of each level based on the fusion weight and the fusion feature maps to be processed at each level; the predicted feature map is the adaptive fusion feature map.

[0126] like Figure 2 and 4 In this process, an adaptive spatial feature fusion module ASFF is set for each of the three feature maps to be processed, namely ASFF_1, ASFF_2, and ASFF_3.

[0127] The core idea of ​​adaptive feature fusion is to adaptively learn the fusion space weights of feature maps at different scales, thereby more fully fusing feature maps containing different information and reducing interference from background information. This embodiment performs adaptive fusion on three feature layers with different resolutions: level_1, level_2, and level_3 (corresponding feature map resolutions of 20×20×512, 40×40×256, and 80×80×128). The predicted feature map on any of the adjusted feature layers is obtained by fusing the feature maps from level_1, level_2, and level_3.

[0128] make This represents the feature vector at position (i, j) in the fused feature map of layer m to be processed, which is then adjusted to the fused feature map of layer n to be processed. These are the fusion weights when the fusion feature maps at three different levels (level_1, level_2, and level_3) are adjusted to the nth level fusion feature map; n = 1, 2, 3; m = 1, 2, 3.

[0129] For any one of level_1, level_2, and level_3 (i.e., level_n), the final predicted feature map after adaptive spatial feature fusion is as follows:

[0130] (7)

[0131] In the formula, This represents the adaptive fusion feature at position (i, j) in the nth layer predicted feature map; ;and Defined using the softmax function :

[0132] (8)

[0133] (9)

[0134] (10)

[0135] in,

[0136] From It is learned adaptively.

[0137] In step S3, an adaptive spatial feature fusion module is introduced into the neck of the feature fusion layer. It utilizes feature information from different receptive fields for adaptive fusion, mitigating interference from background and other conflicting information. Since feature information of small targets is easily lost during downsampling in the convolutional neural network, the feature map output from the feature pyramid is highly susceptible to contamination from background and other conflicting information. Therefore, an adaptive spatial feature fusion module is used after the feature pyramid to adaptively learn the spatial fusion weights of feature maps with three different receptive fields: 80×80×128, 40×40×256, and 20×20×512. This filters out background and other conflicting information and better integrates shrimp larvae feature information from different layers. The re-fused predicted feature maps all contain target information from the three feature maps: 80×80×128, 40×40×256, and 20×20×512.

[0138] S4: The adaptive fusion feature map is input into the improved head regression detection network to obtain the shrimp larvae anchor-free detection result; the improved head regression detection network replaces the traditional cross-union-parallel comparison (CUP) module of the traditional head regression detection network with a generalized CUP module. The application of the improved head regression detection network corresponds to... Figure 2 The regression testing phase in the process.

[0139] Since the traditional Intersection over Union (IoU) ratio cannot accurately reflect the degree of overlap between the predicted bounding box and the actual target location, this invention introduces a more advanced Generalized Intersection over Union (GIoU). GIoU calculates the minimum bounding rectangle of the predicted bounding box and the target ground truth bounding box, applying the non-overlapping portion as a penalty term in the loss calculation. This better reflects the overlap between the predicted and target ground truth bounding boxes. Figure 6 A schematic diagram comparing the traditional cross-union ratio and the generalized cross-union ratio is shown.

[0140] The calculation function expression for the generalized intersection-union-comparison module is as follows:

[0141] (11)

[0142] in, (12)

[0143] In the formula, GIoU represents the generalized intersection-union ratio; A p Represents the predicted bounding box; A g A represents the true bounding box of the target; c Representing A p and A g The minimum bounding rectangle;

[0144] Compared to IoU, GIoU is the lower bound of IoU, with a value range of [-1, 1]. Due to the introduction of penalty, GIoU considers not only overlapping regions but also non-overlapping regions, better reflecting the overlap between the predicted bounding box and the ground truth bounding box. It still effectively optimizes the predicted bounding box even when the two do not overlap.

[0145] The loss function expression for the generalized intersection-union module is:

[0146] (13).

[0147] In step S4, to address the difficulty of locating small targets, the traditional Cross-Union Ratio (CUI) loss is inadequate for optimizing cases where the bounding box and the actual target location do not overlap. A Generalized Cross-Union Ratio (GRU) is used instead of the traditional CUI, applying the non-overlapping portion as a penalty in the loss calculation. This better reflects the overlap between the bounding box and the actual target location. A GRU module is added to the header of the bounding box prediction process to better measure the bounding box regression loss and overcome the problem of the traditional CUI module's inability to optimize cases where the bounding box and the actual target location do not overlap.

[0148] In this embodiment, a decoupled head is used to decouple the classification and regression branches. The coupling conflict between the two sub-tasks of classification and regression is a common problem in object detection. In this method, to resolve the coupling conflict between classification and regression in object detection, the classification and regression branches are decoupled, as follows: Figure 7 As shown, the detection head is decoupled by constructing three parallel branches: one branch is used for target classification, with the same number of output channels as the number of predicted target categories, C; one branch is used for bounding box regression, with 4 output channels, used to predict the position coordinates of the bounding box, i.e., two offsets relative to the top left corner of the grid and the predicted height and width of the box; the third branch is used to calculate the target confidence, with 1 output channel. That is, the improved head regression detection network adopts a decoupled prediction head structure with three parallel branches; the decoupled prediction head structure includes a classification branch for target classification, a regression branch for predicting the bounding box position coordinates, and a confidence branch for calculating the target confidence. Figure 7 As shown, the adaptive fusion feature map is processed by 1×1 Conv_2D and then split into two branches. One branch is processed by 3×3 Conv_2D and 1×1 Conv_2D in sequence to achieve target classification. The other branch is processed by 3×3 Conv_2D and then split into two branches. Both branches are processed by 1×1 Conv_2D to achieve bounding box regression and confidence calculation.

[0149] The decoupled detection head is used to decouple the classification and regression branches, avoiding coupling conflicts between them. The feature map is divided into S×S grids, and only one target location is predicted in each grid to achieve anchorless operation.

[0150] In this example, the target detection model, composed of an improved feature extraction backbone network, an improved neck feature fusion network, and an improved head regression detection network, requires training on data before practical target detection applications. The model is trained using the training set from the shrimp larvae dataset until the loss function fully converges. The weight file with the best accuracy is recorded, and the performance of the method in shrimp larvae target detection is validated on the test set. The results show that the proposed method outperforms Faster RCNN, SSD, CenterNet, FCOS, YOLOv5_s, etc., demonstrating that the proposed method can effectively overcome the influence of shrimp larvae blur and small size on detection, and can be better applied to intensive aquaculture environments.

[0151] This embodiment provides an anchorless detection method for shrimp larvae in aquaculture. Starting from the characteristics of shrimp larvae in actual aquaculture and the difficulties in shrimp larvae detection, this method introduces a coordinate attention module, an adaptive spatial feature fusion module, and a generalized intersection-union (IUU) module in the feature extraction stage (backbone network), the feature fusion stage (neck), and the regression detection stage (head). This enhances the method's feature extraction capability for fuzzy, small-sized shrimp larvae targets, reduces interference from background and other conflicting information, and improves bounding box optimization regression. The proposed method achieves excellent performance in shrimp larvae detection.

[0152] Example 2

[0153] This embodiment provides a system for detecting anchorless shrimp larvae in aquaculture, the system comprising:

[0154] The image acquisition unit T1 is used to acquire images of shrimp larvae to be tested collected on-site from aquaculture bases.

[0155] The feature extraction unit T2 is used to input the shrimp seedling image to be detected into the improved feature extraction backbone network to obtain multiple feature maps with coordinate attention processing; the improved feature extraction backbone network is to connect a coordinate attention mechanism module to the output end of each output branch of the traditional feature extraction backbone network.

[0156] Specifically, the feature extraction unit T2 includes:

[0157] The feature map acquisition subunit T21 is used to input the shrimp seedling image to be detected into the improved feature extraction backbone network to obtain the feature map to be processed generated by the previous layer of each coordinate attention mechanism module.

[0158] The feature encoding subunit T22 is used to encode each channel along the horizontal and vertical directions of each feature map to be processed, respectively, to obtain horizontally encoded features and vertically encoded features.

[0159] The splicing subunit T23 is used to perform splicing operations based on the horizontal direction encoding features and the vertical direction encoding features to obtain an intermediate feature map with horizontal and vertical direction encoding information.

[0160] The splitting subunit T24 is used to split the intermediate feature map into two independent tensors, resulting in a horizontal tensor and a vertical tensor.

[0161] The transformation subunit T25 is used to convert the horizontal tensor and the vertical tensor into tensors with the same number of channels as the feature map to be processed, respectively, to obtain the horizontal target tensor and the vertical target tensor.

[0162] The coordinate attention application subunit T26 is used to apply the horizontal target tensor and the vertical target tensor to each position of each channel of the feature map to be processed, so as to obtain the output feature map processed by the coordinate attention mechanism module; the output feature map is the feature map processed by the coordinate attention.

[0163] The adaptive feature fusion unit T3 is used to input multiple feature maps processed by the coordinate attention process into the improved neck feature fusion network to obtain multiple adaptive fused feature maps; the improved neck feature fusion network is a traditional neck feature fusion network that employs an adaptive spatial feature fusion module; the adaptive spatial feature fusion module and the adaptive fused feature map... Figure 1 One-to-one correspondence.

[0164] Specifically, the adaptive feature fusion unit T3 includes:

[0165] The subunit T31 for obtaining the fusion feature map to be processed is used to input the feature maps of multiple coordinate attention processing into the improved neck feature fusion network to obtain the fusion feature map to be processed generated by the previous layer of each adaptive spatial feature fusion module; each fusion feature map to be processed is a fusion feature map at a different level; the different levels of the fusion feature map to be processed are divided according to different resolutions.

[0166] The adaptive fusion subunit T32 is used to input the fusion feature maps to be processed at each level into each of the adaptive spatial feature fusion modules, adaptively calculate the fusion weight at each position of the fusion feature maps to be processed at each level, and obtain the predicted feature map of each level based on the fusion weight and the fusion feature maps to be processed at each level; the predicted feature map is the adaptive fusion feature map.

[0167] The shrimp larvae anchorless detection unit T4 is used to input the adaptive fusion feature map into the improved head regression detection network to obtain the shrimp larvae anchorless detection result; the improved head regression detection network is to replace the traditional cross-over and union ratio module of the traditional head regression detection network with the generalized cross-over and union ratio module.

[0168] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0169] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for detecting anchorless shrimp larvae in aquaculture, characterized in that, The method includes: Obtain the dataset of shrimp larvae images to be detected; The dataset of shrimp larvae images to be detected is input into an improved feature extraction backbone network to obtain multiple feature maps with coordinate attention processing. The improved feature extraction backbone network is a traditional feature extraction backbone network with a coordinate attention mechanism module connected to the output of each output branch. Multiple feature maps processed by the coordinate attention method are input into an improved neck feature fusion network to obtain multiple adaptive fusion feature maps. The improved neck feature fusion network is an adaptive spatial feature fusion module that is used in a traditional neck feature fusion network. The adaptive spatial feature fusion module corresponds one-to-one with the adaptive fusion feature map. The adaptive fusion feature map is input into the improved head regression detection network to obtain the shrimp larvae anchorless detection result; the improved head regression detection network is to replace the traditional cross-union-over-union module of the traditional head regression detection network with the generalized cross-union-over-union module; The improved feature extraction backbone network comprises three modules connected in sequence; the first module comprises a Focus layer, a first CBS layer, a first CSP layer, a second CBS layer, a second CSP layer, and a first coordinate attention mechanism module connected in sequence; the second module comprises a third CBS layer, a third CSP layer, and a second coordinate attention mechanism module connected in sequence; and the third module comprises a fourth CBS layer, a SPP layer, a fourth CSP layer, and a third coordinate attention mechanism module. The improved neck feature fusion network includes a fourth module, a fifth module, and a sixth module; The fourth module includes the fifth CBS layer, the first UpSample layer, the first splicing layer, the fifth CSP layer, the sixth CBS layer, the second UpSample layer, and the second splicing layer, which are connected in sequence. The fifth module consists of the sixth CSP layer, the seventh CBS layer, the third splicing layer, the seventh CSP layer, the eighth CBS layer, the fourth splicing layer, and the eighth CSP layer connected in sequence. The sixth module includes three adaptive spatial feature fusion modules; The output of the first coordinate attention mechanism module is also connected to the input of the second stitching layer, the output of the second coordinate attention mechanism module is also connected to the input of the first stitching layer, and the output of the third coordinate attention mechanism module is connected to the input of the fifth CBS layer. The output of the fifth CBS layer is also connected to the input of the fourth splicing layer, the output of the sixth CBS layer is also connected to the input of the third splicing layer, and the output of the second splicing layer is connected to the input of the sixth CSP layer. The outputs of the sixth CSP layer to the eighth CSP layer are used as inputs to each adaptive spatial feature fusion module.

2. The method according to claim 1, characterized in that, The dataset of shrimp larvae images to be detected is input into the improved feature extraction backbone network, specifically including: The dataset of shrimp larvae images to be detected is input into the improved feature extraction backbone network to obtain the feature maps to be processed generated by the layer before each coordinate attention mechanism module; For each feature map to be processed, each channel is encoded along the horizontal and vertical directions of the feature map to obtain horizontal encoded features and vertical encoded features. By concatenating the horizontal and vertical coding features, an intermediate feature map with both horizontal and vertical coding information is obtained. The intermediate feature map is split into two independent tensors to obtain a horizontal tensor and a vertical tensor; The horizontal tensor and the vertical tensor are respectively converted into tensors with the same number of channels as the feature map to be processed, to obtain the horizontal target tensor and the vertical target tensor; The horizontal target tensor and the vertical target tensor are applied to each position of each channel of the feature map to be processed to obtain the output feature map processed by the coordinate attention mechanism module; the output feature map is the feature map processed by the coordinate attention mechanism.

3. The method according to claim 2, characterized in that, The expression for the horizontal directional encoding feature is: In the formula, Indicates the first in the horizontal direction c Channel height h The encoded features obtained at the location; W Indicates the width of the feature map to be processed; Indicates the feature map to be processed. c Channel 1 h The first in the row i One eigenvalue; The expression for the vertical direction encoded feature is: In the formula, Indicates the first in the vertical direction c Channel width w The encoded features obtained at the location; H Indicates the height of the feature map to be processed; Indicates the feature map to be processed. c Channel 1 w The first in the column j One eigenvalue; The expression for the intermediate feature map is: in, f It is an intermediate feature map containing encoding information in both the horizontal and vertical directions. z h This indicates a feature map generated after all channels of the feature map to be processed have been encoded with features in the horizontal direction. z w This indicates that the feature map to be processed is generated after all channels of the feature map have been encoded with features in the vertical direction; z h , z w ] indicates that z h and z w Perform spatial dimension concatenation operations; F 1 represents the transformation function implemented using a 1×1 convolution; It is a non-linear activation function; The expression for the horizontal target tensor is: in, g h Represents the horizontal target tensor; This represents the tensor containing horizontally encoded information obtained from the splitting process. F h This represents a transformation function implemented using a 1×1 convolution; Represents the Sigmoid function; The expression for the vertical target tensor is: in, g w Represents the vertical target tensor; This represents the tensor containing vertically encoded information obtained from the splitting process. F w This represents a transformation function implemented using a 1×1 convolution; The expression for the output feature map is: in, y c ( i , j ) represents the feature map to be processed. c Location of the passage ( i , j The new feature value at position ) after processing by the coordinate attention mechanism; x c (i , j ) represents the feature map to be processed. c Location of the passage ( i , j The eigenvalue at () is a position; Indicates the first c Channel height h The first in i Attention weights at each location coordinate; Indicates the first c Channel width w The first in j Attention weights at each location coordinate.

4. The method according to claim 3, characterized in that, The feature maps processed by the coordinate attention method are input into the improved neck feature fusion network to obtain multiple adaptive fusion feature maps, specifically including: The feature maps of the coordinate attention processing are input into the improved neck feature fusion network to obtain the fusion feature map to be processed generated by the previous layer of each adaptive spatial feature fusion module; each fusion feature map to be processed is a fusion feature map at a different level; the different levels of the fusion feature map to be processed are divided according to different resolutions; The fusion feature maps to be processed at each level are input into each of the adaptive spatial feature fusion modules, and the fusion weights at each position of the fusion feature maps to be processed at each level are adaptively calculated; and the predicted feature map of each level is obtained based on the fusion weights and the fusion feature maps to be processed at each level; the predicted feature map is the adaptive fusion feature map.

5. The method according to claim 4, characterized in that, The adaptive fusion expression of the adaptive spatial feature fusion module is as follows: In the formula, Indicates the first n Location in the layer prediction feature map ( i , j Adaptive fusion features at () location; ; Indicates the first m Location of the layer to be processed in the fusion feature map ( i , j Adjusted to the first position. n Feature vectors after the fused feature maps of the layers to be processed; The three different levels of unprocessed fusion feature maps were adjusted to the [number]th [level]. n Fusion weights when processing fused feature maps in layers; n =1，2，3； m =1，2，3。 6. The method according to claim 1, characterized in that, The calculation function expression for the generalized intersection-union-comparison module is as follows: in, ; In the formula, GIoU Indicates generalized intersection, union, and comparison; A p Represents the predicted bounding box; A g Represents the true bounding box of the target; A c represent A p and A g The minimum bounding rectangle; The loss function expression for the generalized intersection-union module is: 。 7. The method according to claim 1, characterized in that, The improved head regression detection network employs a decoupled prediction head structure with three parallel branches. The decoupled prediction head structure includes a classification branch for target classification, a regression branch for predicting bounding box location coordinates, and a confidence branch for calculating target confidence.

8. A non-anchored detection system for shrimp larvae in aquaculture, characterized in that, The system includes: The image acquisition unit is used to acquire images of shrimp larvae to be tested, collected on-site from aquaculture bases. The feature extraction unit is used to input the shrimp seedling image to be detected into the improved feature extraction backbone network to obtain multiple feature maps with coordinate attention processing; the improved feature extraction backbone network is a coordinate attention mechanism module connected to the output end of each output branch of the traditional feature extraction backbone network. An adaptive feature fusion unit is used to input multiple feature maps processed by the coordinate attention into an improved neck feature fusion network to obtain multiple adaptive fusion feature maps; the improved neck feature fusion network is an adaptive spatial feature fusion module used in a traditional neck feature fusion network; the adaptive spatial feature fusion module corresponds one-to-one with the adaptive fusion feature map; Anchorless shrimp larvae detection unit is used to input the adaptive fusion feature map into the improved head regression detection network to obtain anchorless shrimp larvae detection results; the improved head regression detection network replaces the traditional cross-over and union (COU) module of the traditional head regression detection network with a generalized COU module. The improved feature extraction backbone network comprises three modules connected in sequence; the first module comprises a Focus layer, a first CBS layer, a first CSP layer, a second CBS layer, a second CSP layer, and a first coordinate attention mechanism module connected in sequence; the second module comprises a third CBS layer, a third CSP layer, and a second coordinate attention mechanism module connected in sequence; and the third module comprises a fourth CBS layer, a SPP layer, a fourth CSP layer, and a third coordinate attention mechanism module. The improved neck feature fusion network includes a fourth module, a fifth module, and a sixth module; the fourth module includes a fifth CBS layer, a first UpSample layer, a first stitching layer, a fifth CSP layer, a sixth CBS layer, a second UpSample layer, and a second stitching layer connected in sequence; the fifth module includes a sixth CSP layer, a seventh CBS layer, a third stitching layer, a seventh CSP layer, an eighth CBS layer, a fourth stitching layer, and an eighth CSP layer connected in sequence; the sixth module includes three adaptive spatial feature fusion modules; The output of the first coordinate attention mechanism module is also connected to the input of the second stitching layer, the output of the second coordinate attention mechanism module is also connected to the input of the first stitching layer, and the output of the third coordinate attention mechanism module is connected to the input of the fifth CBS layer. The output of the fifth CBS layer is also connected to the input of the fourth splicing layer, the output of the sixth CBS layer is also connected to the input of the third splicing layer, and the output of the second splicing layer is connected to the input of the sixth CSP layer. The outputs of the sixth CSP layer to the eighth CSP layer are used as inputs to each adaptive spatial feature fusion module.

9. The system according to claim 8, characterized in that, The feature extraction unit specifically includes: The feature map acquisition subunit is used to input the shrimp seedling image to be detected into the improved feature extraction backbone network to obtain the feature map to be processed generated by the previous layer of each coordinate attention mechanism module; The feature encoding subunit is used to encode each channel along the horizontal and vertical directions of each feature map to be processed, respectively, to obtain horizontally encoded features and vertically encoded features. The splicing subunit is used to perform splicing operations based on the horizontal direction encoding features and the vertical direction encoding features to obtain an intermediate feature map with horizontal and vertical direction encoding information. The splitting subunit is used to split the intermediate feature map into two independent tensors, resulting in a horizontal tensor and a vertical tensor. The transformation subunit is used to convert the horizontal tensor and the vertical tensor into tensors with the same number of channels as the feature map to be processed, respectively, to obtain the horizontal target tensor and the vertical target tensor; The coordinate attention application subunit is used to apply the horizontal target tensor and the vertical target tensor to each position of each channel of the feature map to be processed, so as to obtain the output feature map processed by the coordinate attention mechanism module; the output feature map is the feature map processed by the coordinate attention.

10. The system according to claim 9, characterized in that, The adaptive feature fusion unit specifically includes: The subunit for obtaining the fused feature map to be processed is used to input the feature maps of multiple coordinate attention processing into the improved neck feature fusion network to obtain the fused feature map to be processed generated by the previous layer of each adaptive spatial feature fusion module; each of the fused feature maps to be processed is a fused feature map at a different level; the different levels of the fused feature map to be processed are divided according to different resolutions; An adaptive fusion subunit is used to input the fusion feature maps to be processed at each level into each of the adaptive spatial feature fusion modules, adaptively calculate the fusion weight at each position of the fusion feature maps to be processed at each level, and obtain the predicted feature map at each level based on the fusion weight and the fusion feature maps to be processed at each level; the predicted feature map is the adaptive fusion feature map.

Images