Sick juvenile fish identification method based on improved YOLOv10 model

Abstract

The invention provides a sick juvenile fish identification method based on an improved YOLOv10 model, and the method comprises the steps: collecting a fish image in a water area as a to-be-detected image, and carrying out the preprocessing of the to-be-detected image, and obtaining an image data set; the method comprises the following steps: constructing a YOLOv10-UDFishNet model; the image data set is input into the YOLOv10-UFFishNet model, the YOLOv10-UFFishNet model is trained through the image data set, and the trained YOLOv10-UFFishNet model is obtained; and inputting a fish image in a water area to be detected into the trained YOLOv10-UFFishNet model to obtain a fish detection result and position information of a fish individual target. According to the method, the sick juvenile fish in the turbid underwater environment can be accurately and efficiently identified.

Description

Technical Field

[0001] The invention belongs to the technical field of intelligent identification, and specifically discloses a method for identifying sick fry based on an improved YOLOv10 model.

[0002] Background

[0003] As a farmed fish with high economic value, the health status and disease prevention of redfin pufferfish directly affect the sustainable development and economic benefits of the aquaculture industry. Therefore, accurate identification of sick juvenile redfin pufferfish is not only the key to ensuring fish health and reducing mortality, but also an important driving force for the growth of aquaculture income. However, due to the particularity of the aquaculture environment, the turbid underwater conditions cause the quality of the acquired images to degrade, which restricts the improvement of the accuracy of diseased juvenile fish detection. In addition, the individual size of redfin pufferfish juveniles is small, which further aggravates the difficulty of diseased juvenile fish detection. Therefore, studying a diseased juvenile fish detection method that can adapt to the turbid underwater environment and effectively deal with redfin pufferfish juveniles is of great significance for improving aquaculture benefits and promoting the sustainable development of the aquaculture industry.

[0004] Computer vision, with its powerful image processing and pattern recognition capabilities, can accurately extract the characteristic information of target objects, making it possible to accurately identify diseased juvenile fish underwater. With the combination of computer vision and deep learning technology, some studies have achieved preliminary judgments on diseased juvenile fish individuals by capturing the appearance changes of fish, but diagnosis based solely on appearance features is often limited. Since fish behavior is a direct reflection of fish health, comprehensive analysis of fish appearance and behavior characteristics can achieve more accurate detection of diseased juvenile fish. Li et al. fused fish contour information with the original image and used an improved YOLOv5 to enhance the position information of fish in the image, which can accurately detect early abnormal behaviors of fish. However, this method relies on a relatively clean data set during training, and its detection effect is still insufficient in the face of complex and changing underwater environments. In order to cope with the challenges of underwater interference in detecting abnormal individuals, Cai et al. designed the NAM-YOLOv7 network by integrating the NAM Attention mechanism, which effectively focused on the key appearance and behavior characteristics of abnormal fish, while discarding irrelevant background information. It has high detection accuracy, but is not suitable for aquaculture environments under strong light. Liu et al. introduced deformable convolution into the YOLOv8 model to enhance the model's behavioral discrimination ability in light fluctuation scenes and achieved significant detection results, but did not consider the adverse effects of other underwater factors on the detection of sick fry. In order to accurately identify sick fry in complex aquaculture environments, Li et al. proposed a model called DDEYOLOv10 based on the YOLOv10 model by introducing the DRNELAN4 module to enhance the model's global feature extraction ability. Although it is beneficial for the detection of sick fry in turbid scenes, the DRNELAN4 module increases the receptive field of the model, resulting in the loss of feature detail information and reducing the detection accuracy of small target sick fry individuals. Wang et al. proposed a sick fry detection model DFYOLO, which combines multiple convolution kernels into a convolution kernel group to enhance the network's ability to extract fish surface features. At the same time, the CBAM module is introduced to suppress the interference of the background on the target, which is beneficial for the precise positioning of small targets in turbid underwater environments and achieves better detection of small target sick fry. However, due to the high similarity between sick individuals and normal fish, this method still has a high rate of missed detection of sick fry.

[0005] In summary, computer vision technology still faces challenges in the underwater diseased fry detection task. The complex underwater environment significantly reduces the image quality, greatly limiting the effective extraction of feature information of diseased fry, resulting in a decrease in the accuracy of diseased fry detection. The small size of diseased fry causes the relatively limited key information they carry, which increases the difficulty of accurately distinguishing diseased fry from normal fish, resulting in a low recall rate for diseased fry detection. In view of the above problems, it is necessary to design a new diseased fry recognition method based on the improved YOLOv10 model to overcome the problems existing in the existing diseased fry recognition method. Summary of the invention

[0006] In order to solve the problems existing in the existing sick fry identification method, such as poor extraction of sick fry feature information in complex water environments and low accuracy of sick fry detection, a sick fry identification method based on an improved YOLOv10 model is proposed.

[0007] The present invention provides a method for identifying diseased fry based on an improved YOLOv10 model, comprising the following steps:

[0008] S1. Collecting fish images in water areas as images to be detected, preprocessing the images to be detected to obtain an image data set;

[0009] S2. construct a YOLOv10-UDFishNet model, including replacing the strided convolutions of Stage 1 and Stage 2 in the backbone network of the YOLOv10 model with SPD-Conv modules, adding auxiliary branches to the backbone network of the YOLOv10 model, adding an HFG-based fusion feature fusion method to the neck network of the YOLOv10 model, and replacing the loss function of the YOLOv10 model with a SlideLoss loss function to obtain the YOLOv10-UDFishNet model;

[0010] S3. Input the image data set obtained after preprocessing in step S1 into the YOLOv10-UDFishNet model obtained in step S2, train the YOLOv10-UDFishNet model with the image data set, and obtain a trained YOLOv10-UDFishNet model;

[0011] S4. Input the fish image in the water area to be detected into the trained YOLOv10-UDFishNet model obtained in step S3 to obtain the fish detection results and the location information of individual fish targets.

[0012] According to a method for identifying sick fry based on an improved YOLOv10 model in some embodiments of the present application, in step S1, the preprocessing includes adjusting the size of the image to be detected, and marking both the appearance abnormalities and the behavior abnormalities in the image to be detected as abnormal labels, and marking normal individuals as normal labels.

[0013] According to a method for identifying sick fry based on an improved YOLOv10 model in some embodiments of the present application, in step S2, the SPD-Conv module includes a space-to-depth layer and a non-step convolution layer, and the space-to-depth layer is used to slice a given feature map X(S, S, C1) to obtain scale 2 The size is The sub-feature maps are concatenated in the channel dimension to obtain a new feature map

[0014] The slicing operation formula is shown in formula (1):

[0015]

[0016] Among them, f a,b represents a sub-feature map, a∈(0,1,2…scale-1), b∈(0,1,2…scale-1), S represents the spatial size of the feature map, C1 represents the number of feature map channels, and scale represents the number of sub-feature maps that are split.

[0017] The non-strided convolution layer is used to add a non-strided convolution layer using C2 convolution kernels to obtain the final feature map X" Where C2 <C1<scale 2 .

[0018] According to a method for identifying sick fry based on an improved YOLOv10 model in some embodiments of the present application, in step S2, the auxiliary branch includes a first auxiliary branch and a second auxiliary branch, the first auxiliary branch is connected in parallel with Stage1, and the second auxiliary branch is connected in parallel with Stage2;

[0019] The first auxiliary branch includes an SPD-Conv module and an improved ShuffleBlock module connected in sequence, and the second auxiliary branch includes an improved ShuffleBlock module; the improved ShuffleBlock module includes an improved feature extraction unit and an improved downsampling unit;

[0020] The improved feature extraction unit replaces the 1×1 convolution after the depth-separable convolution of the feature extraction unit in the basic module ShuffleBlock of ShuffleNetv2 with a partial convolution, wherein the partial convolution performs a standard convolution operation on part of the input channels to extract spatial features, and the remaining channels remain unchanged, and the partial ratio of the partial convolution is r=1 / 2;

[0021] The improved downsampling unit is to replace the depth-separable convolution of the downsampling unit in the basic module ShuffleBlock of ShuffleNetv2 with an SPD-Conv module.

[0022] According to a method for identifying sick fry based on an improved YOLOv10 model in some embodiments of the present application, in step S2, the HFG-based fusion feature fusion method includes the following steps:

[0023] Step a. Given a low-level feature map X low ∈R H×W×C and high-level feature map X high ∈R H×W×C , the high-level feature map X high ∈R H×W×C The channels are evenly split into G groups to obtain G groups of feature maps. If they cannot be evenly divided, they are rounded down.

[0024] Step b. Perform 1×1Conv operation on each group of feature maps after splitting, so that the number of channels of each group of feature maps is the same as that of the low-level feature map X low The number of channels is the same, and each set of feature maps after dimensionality increase is recorded as X g ;

[0025] Step c. Each set of feature maps X g Respectively with X low After adding each element, the feature graphs of each G group are accumulated to obtain the feature graph X′ high , as shown in formula (2):

[0026]

[0027] Among them, g∈(1,2…G);

[0028] Step d. The feature map X′ high Global average pooling is performed in the spatial dimension to compress the entire spatial information into a global feature and obtain a C-dimensional global feature. The compression process is shown in formula (3):

[0029]

[0030] Among them, z c ∈R crepresents the global feature, H represents the height of the feature map, W represents the width of the feature map, and u c represents the input matrix, i represents the row of the input matrix, j represents the column of the input matrix, and c represents the number of channels of the feature map;

[0031] Step e. Pass the global feature through a fully connected layer consisting of a ReLU activation function and a Sigmoid activation function to obtain a weighted output, as shown in formula (4):

[0032] s=σ(W2δ(W1z c )) (4)

[0033] Where s represents the weight output, δ is the ReLU activation function, σ is the Sigmoid activation function, W1(·) represents the fully connected layer operation of the ReLU activation function, and W2(·) represents the fully connected layer operation of the Sigmoid activation function.

[0034] Step f. Combine the weight output with the low-level feature map X low Perform element-by-element multiplication and concatenate with the high-level feature map along the channel to obtain the final output, as shown in formula (5):

[0035]

[0036] Among them, Out represents the final output, represents element-by-element multiplication; [·,·] represents a concatenation operation.

[0037] According to a method for identifying sick fry based on an improved YOLOv10 model in some embodiments of the present application, in step S2, the SlideLoss loss function is as shown in formula (6):

[0038]

[0039] Among them, I n represents the loss of the nth category, n∈(1,2…n), f(x) represents the Slide weighting function added to the baseline model loss function, w n represents the weight factor, y n represents the nth actual value, σ represents the Sigmoid function, x n represents the nth predicted value, and x represents the function input;

[0040] Based on the baseline model loss function, the Slide weighting function f(x) is added as shown in formula (7):

[0041]

[0042] Among them, μ represents the average value of the intersection-over-union ratio between all predicted boxes and the true boxes.

[0043] According to a method for identifying diseased fry based on an improved YOLOv10 model in some embodiments of the present application, in step S3, the improved YOLOv10-UDFishNet model is trained based on the image data set until the training is completed when a set learning round is reached, thereby obtaining a trained YOLOv10-UDFishNet model.

[0044] According to a method for identifying sick fry based on an improved YOLOv10 model in some embodiments of the present application, the fish detection results obtained in step S4 include normal and abnormal.

[0045] The invention proposes a method for identifying sick juvenile fish based on an improved YOLOv10 model, which can accurately and efficiently identify sick juvenile fish in turbid underwater environments. By designing an auxiliary branch based on a backbone network, the interference problem of turbid water background on feature extraction is solved, the feature extraction capability of the network in low-quality images is enhanced, and the detection effect of underwater sick juvenile fish is improved. The SPD-Conv module is used to reduce the loss of small target juvenile fish information caused by the downsampling process. In the neck network, an adaptive fusion method HFG-based fusion feature fusion method guided by high-level features is proposed to realize the efficient reuse of low-level sick juvenile fish features of the network and improve the resolution capability of small target sick juvenile fish. The SlideLoss loss function is introduced to adaptively adjust the sample weights, increase the attention paid to sick juvenile fish samples during training, and further improve the detection performance of sick juvenile fish. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic diagram of a flow chart of a method for identifying sick fry based on an improved YOLOv10 model of the present invention;

[0047] Figure 2 Schematic diagram of the structure of the YOLOv10-UDFishNet model in Example 2 of the present invention;

[0048] Figure 3 Schematic diagram of the structure of the SPD-Conv module in Example 2 of the present invention;

[0049] Figure 4 The figure is a visualization diagram of the feature graphs before and after the backbone network Stage 2 in Example 2 of the present invention, (a) is a visualization diagram of the feature graph before Stage 2, and (b) is a visualization diagram of the feature graph after Stage 2;

[0050] Figure 5(a) is a schematic diagram of the structure of the feature extraction unit of the ShuffleBlock in Example 3 of the present invention, (b) is a schematic diagram of the structure of the downsampling unit of the ShuffleBlock in Example 3 of the present invention, (c) is a schematic diagram of the structure of the improved feature extraction unit of the improved ShuffleBlock module in Example 3 of the present invention, (d) is a schematic diagram of the structure of the improved downsampling unit of the improved ShuffleBlock module in Example 3 of the present invention;

[0051] Figure 6 Schematic diagram of the structure of partial convolution in Embodiment 2 of the present invention;

[0052] Figure 7 Schematic diagram of the structure of the HFG-based fusion feature fusion method in Example 2 of the present invention;

[0053] Figure 8 These are heat maps of the detection effects of different models on small target diseased fry in Example 3 of the present invention, (a) is a heat map of the detection effect of Original image, (b) is a heat map of the detection effect of RT-DETR, (c) is a heat map of the detection effect of YOLOv8, (d) is a heat map of the detection effect of YOLOv9, (e) is a heat map of the detection effect of YOLOv10, (f) is a heat map of the detection effect of YOLOv8-Head-ECAM, (g) is a heat map of the detection effect of DDEYOLOv9, (h) is a heat map of the detection effect of DFYOLO, and (i) is a heat map of the detection effect of YOLOv10-UDFishNet. DETAILED DESCRIPTION

[0054] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0055] Example 1: This example provides a method for identifying sick fry based on an improved YOLOv10 model. Figure 1 As shown, the following steps are included:

[0056] S1. Collect images of fish in water areas as images to be detected, and preprocess the images to be detected to obtain an image data set;

[0057] S2. Build a YOLOv10-UDFishNet model, including replacing the strided convolutions of Stage 1 and Stage 2 in the backbone network of the YOLOv10 model with SPD-Conv modules, adding auxiliary branches to the backbone network of the YOLOv10 model, adding the HFG-based fusion feature fusion method to the neck network of the YOLOv10 model, and replacing the loss function of the YOLOv10 model with the SlideLoss loss function to obtain the YOLOv10-UDFishNet model;

[0058] S3. Input the image dataset obtained after preprocessing in step S1 into the YOLOv10-UDFishNet model obtained in step S2, train the YOLOv10-UDFishNet model with the image dataset, and obtain a trained YOLOv10-UDFishNet model;

[0059] S4. Input the fish image in the water area to be detected into the trained YOLOv10-UDFishNet model obtained in step S3 to obtain the fish detection results and the location information of individual fish targets.

[0060] Example 2: This example provides a method for identifying diseased fry based on an improved YOLOv10 model.

[0061] Due to the turbid underwater environment and the small size of the fry, it is difficult to detect the sick fry of redfin pufferfish. In addition, in order to ensure the real-time detection of sick fry, so as to quickly adjust the corresponding breeding strategy, a high recognition speed is essential for the model. Therefore, the selected model must have both powerful feature extraction capabilities and high real-time response. YOLOv10 is the latest version of the YOLO series algorithm, which exceeds the previous YOLO version and other mainstream models of the same period in terms of the trade-off between accuracy and detection speed. In the YOLOv10 series, YOLOv10x introduces more convolutional layers and residual connections, showing the highest detection accuracy. Therefore, YOLOv10x is used as the baseline model in this embodiment.

[0062] The YOLOv10 model consists of three parts: the input end, the backbone network, and the head network. The input end unifies the image into an image with a resolution of 640×640; the backbone network Backbone consists of a convolutional neural network for extracting image features, and obtains richer gradient flows through the C2f and C2fCIB structures to promote the extraction and learning of richer image features; the head network includes two parts, the neck and the detection head Detect, which are used for feature fusion and target prediction respectively. Among them, the neck is the middle layer connecting the backbone network and the detection head. Through feature fusion and resolution adjustment strategies, the prediction performance of the detection head is significantly improved; the detection head t integrates classifiers and regressors to predict the location and category of the target based on the output features of the neck.

[0063] YOLOv10 is an advanced real-time target detection algorithm that performs well in most scenarios. However, the turbid underwater environment severely restricts the image quality, which directly reduces the detection accuracy of YOLOv10 in aquaculture environments. In addition, the effective information carried by individual juvenile fish in the image is relatively limited, which increases the difficulty of detecting sick juvenile fish. In view of the above problems, this embodiment proposes an improved YOLOv10 model to achieve accurate identification of sick juvenile fish. Its network structure is as follows: Figure 2 As shown in the figure. First, the SPD-Conv module is introduced to optimize the downsampling process and reduce the loss of detail information of larval fish. Secondly, based on the backbone network of YOLOv10, by designing and integrating auxiliary branches, the feature extraction capability of diseased larval fish is enhanced, the diversity and comprehensiveness of feature information are achieved, and the detection effect of diseased larval fish in turbid water conditions is improved. At the same time, a new HFG-based fusion feature fusion method is proposed in the neck network. Through the guidance of high-level semantic information, the channel weights of the low-level feature map are dynamically adjusted to achieve efficient reuse of low-level features of diseased larval fish and adaptive fusion of high- and low-level features, thereby improving the recall rate of small target diseased larval fish. Finally, the SlideLoss loss function is introduced to focus on diseased larval fish samples, which further improves the network's detection performance for underwater diseased larval fish.

[0064] This embodiment provides a method for identifying sick fry based on an improved YOLOv10 model, comprising the following steps:

[0065] S1. Collect images of fish in water areas as images to be detected, and preprocess the images to be detected to obtain an image data set;

[0066] Specifically, the preprocessing includes adjusting the size of the image to be detected, marking the abnormal appearance and abnormal behavior in the image to be detected as abnormal labels, and marking the normal individuals as normal labels; the final marking information can be saved as a .txt format file;

[0067] S2. Build a YOLOv10-UDFishNet model, including replacing the strided convolutions of Stage 1 and Stage 2 in the backbone network of the YOLOv10 model with SPD-Conv modules, adding auxiliary branches to the backbone network of the YOLOv10 model, adding the HFG-based fusion feature fusion method to the neck network of the YOLOv10 model, and replacing the loss function of the YOLOv10 model with the SlideLoss loss function to obtain the YOLOv10-UDFishNet model;

[0068] In the images obtained in the aquaculture environment, the size of the young fish is relatively fixed and relatively small. Directly applying the convolution operation with a stride of 2 as the downsampling method will inevitably lead to a significant loss of fish detail features, increasing the difficulty of detecting small underwater target fish. Therefore, this embodiment uses the SPD-Conv module to replace the original strided convolution in the backbone network Stage1 and Stage2 to reduce the problem of young fish feature loss caused by the downsampling operation.

[0069] In this embodiment, the SPD-Conv module includes a space-to-depth layer and a non-strided convolution layer, such as Figure 3 As shown, the space-to-depth layer is used to slice the given feature map X(S,S,C1) to obtain scale 2 The size is The sub-feature maps are concatenated in the channel dimension to obtain a new feature map

[0070] The slicing operation formula is shown in formula (1):

[0071]

[0072] Among them, f a,b represents a sub-feature map, a∈(0,1,2…scale-1), b∈(0,1,2…scale-1), S represents the spatial size of the feature map, C1 represents the number of feature map channels, and scale represents the number of sub-feature maps that are split.

[0073] The non-strided convolution layer is used to add a non-strided convolution layer using C2 convolution kernels to obtain the final feature map Where C2 <C1<scale 2 .

[0074] To further verify the effect of the optimized SPD-Conv module on small target fish detection, the feature maps before and after the backbone network Stage2 are visualized. Figure 4As shown in the figure, the small target details extracted in Stage 1 and Stage 2 are richer, and these fine-grained information is crucial for the detection of small target individuals; while more abstract and advanced features are extracted after Stage 2, too much fine-grained information may no longer be necessary, and may even introduce noise and redundancy to interfere with the detection performance. Therefore, only the strided convolution in Stage 1 and Stage 2 in the backbone network is replaced, while retaining the fine-grained information, avoiding interference to the subsequent stages, thereby improving the network's detection effect on small target fish.

[0075] Specifically, due to the limited feeding activity of young fish, excessive feed often remains in the water, causing turbidity in the aquaculture water and increasing the difficulty of disease detection. To meet this challenge, existing research mainly adopts a composite backbone network structure, which performs feature fusion after a parallel dual backbone network to enhance the diversity of model feature extraction. However, this approach will cause some features to be extracted repeatedly, increasing the redundancy of the overall features. In addition, the dual-backbone design will significantly increase the number of model parameters, which is not conducive to underwater real-time tasks. Therefore, this embodiment is based on the backbone network of YOLOv10, and compensates for the shortcomings of feature extraction of a single backbone network by designing and integrating auxiliary branches. The auxiliary branches are mainly composed of the basic modules ShuffleBlock of ShuffleNetv2 stacked together, which ensures that the performance of the model is improved while not significantly increasing the computational burden. Figure 2 As shown on the left, the input feature channel is first split, and feature extraction is performed through the backbone network and auxiliary branches respectively. After Stage 1 and Stage 2, the channels are spliced ​​together to achieve diversity and richness in feature extraction. To ensure that the feature map size is consistent when the channels are spliced, a downsampling operation is added before the first ShuffleBlock, thereby ensuring the smooth progress of the subsequent feature fusion process. ShuffleBlock contains two basic units, namely, a feature extraction unit with a convolution stride of 1 (stride = 1) and a downsampling unit with a convolution stride of 2 (stride = 2). The unit structures are shown as follows: Figure 5 (a) Figure 5 (b) However, the traditional ShuffleBlock structure needs to maintain the same number of input and output channels, which limits the improvement of model efficiency to a certain extent. Therefore, the structure of ShuffleBlock is optimized.

[0076] Therefore, in this embodiment, the auxiliary branch includes a first auxiliary branch and a second auxiliary branch, the first auxiliary branch is connected in parallel with Stage1, and the second auxiliary branch is connected in parallel with Stage2;

[0077] The first auxiliary branch includes an SPD-Conv module and an improved ShuffleBlock module connected in sequence, and the second auxiliary branch includes an improved ShuffleBlock module; the improved ShuffleBlock module includes an improved feature extraction unit and an improved downsampling unit;

[0078] like Figure 5 As shown in (c), the improved feature extraction unit replaces the 1×1 convolution after the depthwise separable convolution of the feature extraction unit in the basic module ShuffleBlock of ShuffleNetv2 with partial convolution. The partial convolution performs standard convolution operations on some input channels to extract spatial features, and the remaining channels remain unchanged. This feature significantly reduces the amount of memory access. Its structure is as follows Figure 6 As shown, for the feature map X∈R h×w×c , take the continuous c p The channels are used as representatives of the entire feature map for convolution calculation. At this time, the FLOPs of the partial convolution PConv is only h×w×k 2 ×c p 2 , while the FLOPs of conventional convolution is h×w×k 2 ×c 2 In order to make the model learn more refined feature representation, this embodiment sets the partial ratio of partial convolution to r=1 / 2, so that more input channels can participate in the calculation while still saving a large amount of computing cost. In addition, although some channels are not directly involved in the convolution calculation, the improved channel shuffle mechanism in ShuffleBlock effectively compensates for this potential information loss. By disrupting the order of channels, the information between different channels can fully interact and fuse, thereby providing comprehensive information from all channels for the partial convolution PConv operation in the next stage. This mechanism ensures that even if some channels are not directly involved in the calculation at a specific stage, their information can still be effectively used in subsequent processing, thereby reducing the risk of information loss and improving the overall performance of the model;

[0079] like Figure 5 As shown in (d), the improved downsampling unit is to replace the depth-separable convolution of the downsampling unit in the basic module ShuffleBlock of ShuffleNetv2 with the SPD-Conv module, so as to achieve dual optimization of the backbone network and the auxiliary branch in terms of feature preservation;

[0080] In a complex underwater environment, the feature boundary between diseased fry and healthy fish becomes blurred due to their high similarity. At the same time, the effective feature information carried by the fry in the image is relatively limited, which directly leads to a decrease in the recall rate of small target diseased fry. Therefore, making full use of the rich local semantic information in the low-level features is crucial for accurately identifying small target diseased fry. However, traditional methods often simply splice high and low-level features along the channel and cannot dynamically adjust the importance of different channels of the low-level feature map. Therefore, this embodiment designs a HFG-based fusion feature fusion method in the neck network, such as Figure 7 As shown in the figure, high-level semantic information is used to guide the learning of low-level feature map channel weights, and high-level and low-level features are adaptively fused, so as to efficiently reuse the low-level features of diseased fry and achieve a more accurate individual state distinction effect. Specifically, the HFG-based fusion feature fusion method includes the following steps:

[0081] Step a. Given a low-level feature map X low ∈R H×W×C and high-level feature map X high ∈R H×W×C , the high-level feature map X high ∈R H×W×C The channels are evenly split into G groups. In this embodiment, G=4 to obtain G groups of feature maps. When the channels cannot be evenly divided, they are rounded down.

[0082] Step b. Perform 1×1Conv operation on each group of feature maps after splitting, so that the number of channels of each group of feature maps is the same as that of the low-level feature map X low The number of channels is the same, and each set of feature maps after dimensionality increase is recorded as X g ;

[0083] Step c. Each set of feature maps X g Respectively with X low After adding each element, the feature graphs of each G group are accumulated to obtain the feature graph X′ high , as shown in formula (2):

[0084]

[0085] Among them, g∈(1,2…G);

[0086] Step d. Feature map X′ high Global average pooling is performed in the spatial dimension to compress the entire spatial information into a global feature and obtain a C-dimensional global feature. The compression process is shown in formula (3):

[0087]

[0088] Among them, z c ∈R crepresents the global feature, H represents the height of the feature map, W represents the width of the feature map, and u c represents the input matrix, i represents the row of the input matrix, j represents the column of the input matrix, and c represents the number of channels of the feature map;

[0089] Step e. Pass the global features through a fully connected layer consisting of a ReLU activation function and a Sigmoid activation function to automatically learn the feature importance of each channel and obtain the weighted output, as shown in formula (4):

[0090] s=σ(W2δ(W1z c )) (4)

[0091] Where s represents the weight output, δ is the ReLU activation function, σ is the Sigmoid activation function, W1(·) represents the fully connected layer operation of the ReLU activation function, and W2(·) represents the fully connected layer operation of the Sigmoid activation function.

[0092] Step f. Combine the weight output with the low-level feature map X low Perform element-by-element multiplication to achieve adaptive weighting of the low-level feature map, and concatenate it with the high-level feature map along the channel to obtain the final output, as shown in formula (5):

[0093]

[0094] Among them, Out represents the final output, represents element-by-element multiplication; [·,·] represents a concatenation operation.

[0095] In real aquaculture, diseased juvenile fish samples are severely scarce. Through statistical analysis of the images of the self-built dataset, it is found that the ratio of the total number of fish to the number of diseased juvenile fish is about 30:1. This significant imbalance in quantity poses a severe challenge to the training of the network, resulting in insufficient attention to the scarce diseased juvenile fish samples. Therefore, the SlideLoss loss function is introduced as a new loss function to enhance the network's learning ability for diseased juvenile fish samples and further improve the network's detection performance for diseased juvenile fish. Slide is a sample balance function with adaptive parameters, and its calculation formula is shown in formula (7). Slide divides positive and negative samples according to the threshold μ, and sets a higher weight near μ to increase the relative loss of difficult-to-classify samples by adaptively learning the threshold parameters, thereby focusing more attention on the learning of difficult-to-classify samples. The Slide weighting function f(x) is added to the baseline model loss function. Specifically, the SlideLoss loss function is shown in formula (6):

[0096]

[0097] Among them, In represents the loss of the nth category, n∈(1,2…n), f(x) represents the Slide weighting function added to the baseline model loss function, w n represents the weight factor, y n represents the nth actual value, σ represents the Sigmoid function, x n represents the nth predicted value, and x represents the function input;

[0098] Based on the baseline model loss function, the Slide weighting function f(x) is added as shown in formula (7):

[0099]

[0100] Among them, μ represents the average value of the intersection-over-union ratio between all predicted boxes and the true boxes;

[0101] S3. Input the image dataset obtained after preprocessing in step S1 into the YOLOv10-UDFishNet model obtained in step S2, train the YOLOv10-UDFishNet model with the image dataset, and obtain a trained YOLOv10-UDFishNet model;

[0102] Specifically, the improved YOLOv10-UDFishNet model is trained based on the image data set until the training is completed when the set learning round is reached, and the trained YOLOv10-UDFishNet model is obtained;

[0103] S4. Input the fish image in the water area to be detected into the trained YOLOv10-UDFishNet model obtained in step S3 to obtain the fish detection results and the location information of the individual fish targets;

[0104] Specifically, the obtained fish test results include normal and abnormal.

[0105] Embodiment 3, this embodiment selects red fin pufferfish fry as the experimental object, adopts the sick fry identification method based on the improved YOLOv10 model in embodiment 2, and the sick fry data set is indispensable for studying the detection of sick fry. In view of the lack of public sick fry data sets, it is necessary to carry out the collection of sick fry data sets. In order to ensure the representativeness and authenticity of the data set, field investigations were conducted in the farm many times before data collection, and based on the investigation results, the data collection plan was planned. Afterwards, in the laboratory, the farm environment simulation experiment was carried out by controlling relevant parameters. Finally, after arranging the data set obtained from the simulated environment and the data set obtained from the real breeding pond, a sick fry data set was constructed.

[0106] In this embodiment, red fin pufferfish fry were selected as experimental subjects, with a body length of 7-10 cm, a body weight of 15-18 grams, and a total of 90 fish. To simulate the breeding conditions in the real environment, after injecting an appropriate amount of water into the water tank and letting it stand for 10 minutes, the experimental environment parameters were measured: the water temperature was maintained at 13-18 ° C, the pH value was 6.5-6.9, and the dissolved oxygen content was not less than 5 mg / L. Experimental layout setting: equipped with LED lights as light sources, installed 1 meter above the inside of the water tank to ensure sufficient and uniform lighting. The camera was set up next to the breeding water tank 20 cm above the water surface, and the other end was connected to the computer and automatically stored in MP4 video file format. In order to avoid the red fin pufferfish from having a stress response to the new environment and affecting the accuracy of the experimental data, the fry were placed in an independent container for adaptive breeding for 2-3 days before conducting subsequent experiments. Data collection lasted for about 40 days, and was collected at two different time periods every day, namely 9-10 am and 16-17 pm. During the data collection process, the camera remained on and connected to the computer in real time to ensure that the video data was stored in the computer hard disk without omission. Subsequently, the collected video files were edited and framed to construct a dataset of diseased juvenile redfin pufferfish.

[0107] Through pathological analysis, it was found that the main disease of redfin pufferfish was infection with Vibrio harveyi, which caused abnormalities in appearance and behavior. The appearance of sick redfin pufferfish was: obvious ulceration on the body surface, forming round spots, exposed muscles, and varying degrees of damage to the fins and tail; abnormal behavior was manifested as: swimming sideways, the fish body was perpendicular to the water surface, etc.

[0108] After obtaining the image of sick juvenile fish, it is necessary to draw a bounding box and assign a label to the target in the image, but the process of selecting the target is extremely cumbersome, and the selection results are subjectively different due to different understanding and judgment of the subject. In this embodiment, data annotation consists of semi-automatic annotation and manual fine-tuning to optimize the data annotation process and improve the annotation quality and efficiency. First, the YOLOv10 model is trained using 1000 manually annotated images of sick juvenile fish, and the label is unified as "fish". After that, the remaining 2000 unlabeled images in this experiment are input into the trained YOLOv10 model for drawing the bounding box. At this time, it is necessary to manually retrieve the sick juvenile fish individuals in 3000 images. Reassign labels to all detected fish, uniformly mark abnormal appearance and abnormal behavior as "diseased", change the remaining normal individuals to the "normal" label, and finally save the annotation information as a .txt format file.

[0109] In order to achieve better training results for the network, the labeled diseased juvenile fish dataset was randomly divided in a ratio of 7:2:1, resulting in 2100, 600, and 300 images for the training set, validation set, and test set, respectively. In order to make the evaluation results of the test set more reliable, the test set and the training set were set to come from different video segments during the dataset division process, effectively avoiding data leakage.

[0110] YOLOv10 is an advanced real-time target detection algorithm that performs well in most scenarios. However, the turbid underwater environment severely restricts the image quality, which directly reduces the detection accuracy of YOLOv10 in aquaculture environments. In addition, the effective information carried by individual juvenile fish in the image is relatively limited, which increases the difficulty of detecting sick juvenile fish. In view of the above problems, this embodiment proposes an improved YOLOv10 model to achieve accurate identification of sick juvenile fish. Its network structure is as follows: Figure 2 As shown in the figure. First, the SPD-Conv module is introduced to optimize the downsampling process and reduce the loss of detail information of larval fish. Secondly, based on the backbone network of YOLOv10, by designing and integrating auxiliary branches, the feature extraction capability of diseased larval fish is enhanced, the diversity and comprehensiveness of feature information are achieved, and the detection effect of diseased larval fish in turbid water conditions is improved. At the same time, a new HFG-based fusion feature fusion method is proposed in the neck network. Through the guidance of high-level semantic information, the channel weights of the low-level feature map are dynamically adjusted to achieve efficient reuse of low-level features of diseased larval fish and adaptive fusion of high- and low-level features, thereby improving the recall rate of small target diseased larval fish. Finally, the SlideLoss loss function is introduced to focus on diseased larval fish samples, which further improves the network's detection performance for underwater diseased larval fish.

[0111] In order to ensure the accuracy and reliability of the experimental data, the training and testing of the underwater diseased larvae detection network in this study were run in the same environment. The experimental environment is shown in Table 1. Experimental hyperparameter settings: batch size is 8, training cycle (epochs) is 300 rounds, and the initial learning rate (learning rate) is 0.0001.

[0112] Table 1 Experimental environment

[0113]

[0114] Recognition accuracy and high inference speed play an important role in the real-time monitoring of diseased juvenile fish underwater. In order to comprehensively evaluate the performance of the proposed method, the confusion matrix is ​​used as an evaluation tool, including the following key indicators: True Positive (TP) represents samples correctly identified as abnormal behaviors, False Positive (FP) represents samples incorrectly identified as abnormal behaviors, and False Negative (FN) represents samples mistakenly identified as background.

[0115] In terms of network detection accuracy, the precision rate (P), recall rate (R) and mean average precision (mAP@0.5) are selected as evaluation indicators. mAP combines the accuracy performance under different categories. Generally speaking, the larger the mAP value, the better the overall performance of the model. In addition, diseased fry detection is a task with specific requirements. Compared with the total recall rate, the recall rate of diseased fry should be the primary indicator for evaluating the detection accuracy of the model. In terms of detection speed, the model parameter amount (Params) and inference speed (Inference speed) are selected as evaluation indicators of model size and real-time performance, respectively.

[0116] The calculation of accuracy P is shown in formula (8):

[0117]

[0118] The calculation of recall rate R is shown in formula (9):

[0119]

[0120] The average mean accuracy mAP is shown in formula (10):

[0121]

[0122] Where N represents the number of categories, n∈(1,2,3,…N), AP represents the average precision of a category,

[0123] AP is shown in formula (11):

[0124]

[0125] In order to verify the effectiveness of the improved part of the YOLOv10-UDFishNet model, an ablation experiment was designed. The experimental scheme is as follows: the auxiliary branch, SPD-Conv module, HFG-based fusion feature fusion method and SlideLoss loss function are added to the baseline model in turn, and YOLOv10-A with auxiliary branch added, YOLOv10-AS with auxiliary branch and SPD-Conv module added, and YOLOv10-ASH with auxiliary branch, SPD-Conv module and HFG-based fusion feature fusion method added are obtained. The experimental results are shown in Table 2.

[0126] Table 2 Ablation experiment results

[0127]

[0128] As shown in Table 2, the results show that the above improvements have certain performance improvements over the baseline model. The final YOLOv10-UDFishNet model achieved 94.0% accuracy and 90.1% recall of diseased fry, respectively, which is a significant improvement of 2.2% and 4.8% compared to the baseline model.

[0129] The complex underwater environment leads to image quality degradation, which makes it difficult to extract model features and directly restricts the accuracy and recall rate of diseased fry detection. By integrating auxiliary branches into the backbone network, not only the comprehensiveness of the model for diseased fry feature extraction is enhanced, but also the diversity complementarity of information is realized, which is particularly critical for the detection of diseased fry in low-quality images, and the accuracy and recall rate of diseased fry are increased by 1.3% and 2.3% respectively, while the number of parameters is reduced and the detection speed is accelerated. This is because the input is split and part of it is guided to the auxiliary branch, which is equivalent to lightweight processing, effectively reducing the number of network parameters; at the same time, the partial convolution PConv in the ShuffleBlock module is improved to perform convolution operations on only some channels of the input feature map, and the other channels maintain the same mapping, thereby reducing computational redundancy and significantly improving computational efficiency. The introduction of the SPD-Conv module can reduce the loss of small target fish detail information in the network deepening, achieve more accurate positioning of underwater fry individuals, and improve the detection accuracy to a certain extent. Afterwards, the HFG-basedfusion feature fusion method adaptively enhances the importance of low-level feature map channels, realizes efficient reuse of low-level diseased fry detail features, promotes the flow of information from shallow to deep layers, and is conducive to distinguishing diseased fry from normal fish individuals, significantly improving the recall rate of diseased fry by 1.8%. The SlideLoss loss function is used to further improve the model's detection performance for diseased fry by weighting samples. Compared with the baseline model, the final network shows significant advantages in both accuracy and speed of diseased fry detection.

[0130] To further verify the effectiveness and advancement of the YOLOv10-UDFishNet model of this embodiment in detecting underwater diseased juvenile fish, models with excellent detection performance in this field were selected for comparative experiments, including mainstream target detection models RT-DETR, YOLOv8, YOLOv9, YOLOv10, YOLOv8-Head-ECAM, and models DDEYOLOv9 and DFYOLO with significant detection effects on underwater diseased juvenile fish. The experimental results are shown in Table 3.

[0131] Table 3 Comparison of detection performance of different models

[0132]

[0133] As can be seen from Table 3, compared with other models, the YOLOv10-UDFishNet model proposed in this embodiment has achieved the best results in terms of comprehensive performance. Compared with RT-DETR, the YOLOv10-UDFishNet model significantly improved the detection accuracy and recall rate of sick fry by 8.3% and 13.7%. This is because in the turbid underwater scene, the global receptive field of the RT-DETR model cannot accurately focus on the sick fry individuals, resulting in serious lack of feature extraction of sick fry, resulting in poor final detection effect. Compared with the mainstream YOLO series algorithms, the YOLOv10-UDFishNet model has achieved significant improvements of 5.6%, 6.0% and 4.8% in the recall rate of sick fry, respectively. This is mainly attributed to the degradation of image quality caused by the turbid underwater environment, blurring the feature boundary between sick fry and normal fish, and the inevitable loss of spatial information in the traditional downsampling process, which weakens the detection ability of small target individuals and affects the detection performance of the model for sick fry. YOLOv8-Head-ECAM effectively alleviates the problem of small target positioning by adding a large-size detection head. Although it shows higher accuracy than the YOLO algorithm, the interference of turbid environment on the detection of sick fry is difficult to avoid, resulting in the recall rate of sick fry failing to surpass the YOLOv10-UDFishNet model of this embodiment.

[0134] Compared with the underwater sick larvae detection models DDEYOLOv9 and DFYOLO, the YOLOv10-UDFishNet model also shows its advancedness. Although DDEYOLOv9 has improved the feature extraction capability of sick larvae to a certain extent by introducing the DRNELAN4 module, its enlarged receptive field ignores the detailed features of small target fish, limiting the improvement in the recall rate; while DFYOLO enhances the model's focus on key image areas and the extraction of fish surface features through the combination of convolution kernel groups and CBAM modules. However, the redfin pufferfish larvae are small in size, and their sick larvae are easily confused with background information, resulting in insufficient improvement in the recall rate of sick larvae. Figure 8The heat map intuitively shows the detection effect of each model on small target diseased juvenile fish. In the figure, (a) is the heat map of the detection effect of the original image, (b) is the heat map of the detection effect of RT-DETR, (c) is the heat map of the detection effect of YOLOv8, (d) is the heat map of the detection effect of YOLOv9, (e) is the heat map of the detection effect of YOLOv10, (f) is the heat map of the detection effect of YOLOv8-Head-ECAM, (g) is the heat map of the detection effect of DDEYOLOv9, (h) is the heat map of the detection effect of DFYOLO, (i) is the heat map of the detection effect of YOLOv10-UDFishNet, and the color depth in the heat map directly reflects the degree of attention paid to the model in the area. Through the color contrast of the heat map, it can be seen that the YOLOv10-UDFishNet model of this embodiment has a high degree of focus and accurate detection ability on small target diseased juvenile fish in an underwater turbid environment.

[0135] In order to evaluate the applicability of the proposed YOLOv10-UDFishNet model and the robustness of small target fish detection, this embodiment further extends the method to the field of target detection and focuses on the challenging small target fish detection task. Specifically, 2420 images are randomly selected from the fish dataset of the China Agricultural Artificial Intelligence Innovation and Entrepreneurship Competition, and the dataset is divided according to the ratio of 7:2:1. The target detection experiment is carried out simultaneously with RT-DETR, YOLOv8-Head-ECAM, YOLOv8, YOLOv9 and the baseline model YOLOv10 to compare the model performance. The experimental results are shown in Table 4:

[0136] Table 4 Comparison of model detection performance on public datasets

[0137]

[0138] As shown in Table 4, the data shows that the YOLOv10-UDFishNet model of this embodiment shows a significant advantage in the recall rate index, which is 8.6% and 1.3% higher than RT-DETR and the baseline model. In addition, compared with the YOLOv8-Head-ECAM integrated with the efficient attention mechanism ECAM, the YOLOv10-UDFishNet model of this embodiment has achieved a consistent level of accuracy in individual fish detection, with an accuracy rate of up to 99.1%. The YOLOv10-UDFishNet model of this embodiment maintains excellent performance in the recall rate and accuracy of fish detection, which fully proves that the YOLOv10-UDFishNet model has a certain robustness in the task of detecting small target diseased juvenile fish underwater.

[0139] In the task of detecting sick juvenile redfin pufferfish in a complex underwater environment, this embodiment proposes a new method for identifying sick juvenile redfin pufferfish based on an improved YOLOv10 model, aiming to accurately and efficiently detect sick juvenile redfin pufferfish in a turbid underwater environment. By designing an auxiliary branch based on the backbone network, the problem of interference of turbid water background on feature extraction is solved, and the adaptability of the model in a complex environment is significantly enhanced. In order to more accurately distinguish sick juvenile fish from normal fish individuals, the SPD-Conv module is used to replace the traditional stride convolution to reduce the loss of small target information caused by the downsampling process; in the neck network, an adaptive fusion method HFG-based fusion feature fusion method based on high-level feature guidance is proposed to achieve efficient reuse of low-level sick juvenile fish features by the network and improve the resolution of small target sick juvenile fish. Finally, the SlideLoss loss function is introduced to adaptively adjust the sample weights, increase the attention to difficult samples, and further improve the network's detection performance for sick juvenile fish. Experimental results show that compared with the baseline model, the improved YOLOv10-UDFishNet model has improved accuracy, total recall, recall rate of diseased fry and average precision by 2.2%, 2.9%, 4.8% and 1.5% respectively.

[0140] The embodiments of the present invention are given for the purpose of illustration and description, and are not intended to be exhaustive or to limit the invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.

Claims

1. A method for identifying sick fry based on an improved YOLOv10 model, characterized in that: The steps include: S1. Collecting fish images in water areas as images to be detected, preprocessing the images to be detected to obtain an image data set; S2. construct a YOLOv10-UDFishNet model, including replacing the strided convolutions of Stage 1 and Stage 2 in the backbone network of the YOLOv10 model with SPD-Conv modules, adding auxiliary branches to the backbone network of the YOLOv10 model, adding an HFG-based fusion feature fusion method to the neck network of the YOLOv10 model, and replacing the loss function of the YOLOv10 model with a SlideLoss loss function to obtain the YOLOv10-UDFishNet model; S3. Input the image data set obtained after preprocessing in step S1 into the YOLOv10-UDFishNet model obtained in step S2, train the YOLOv10-UDFishNet model with the image data set, and obtain a trained YOLOv10-UDFishNet model; S4. Input the fish image in the water area to be detected into the trained YOLOv10-UDFishNet model obtained in step S3 to obtain the fish detection results and the location information of individual fish targets.

2. The method for identifying sick fry based on the improved YOLOv10 model according to claim 1, characterized in that: In the step S1, the preprocessing includes adjusting the size of the image to be detected, marking the abnormal appearance and abnormal behavior in the image to be detected as abnormal labels, and marking the normal individuals as normal labels.

3. The method for identifying sick fry based on the improved YOLOv10 model according to claim 1, characterized in that: In step S2, the SPD-Conv module includes a space-to-depth layer and a non-step convolution layer. The space-to-depth layer is used to slice the given feature map X(S, S, C1) to obtain scale 2 The size is ( C1), the sub-feature maps are concatenated in the channel dimension to obtain a new feature map X′( C1×scale 2 ); The slicing operation formula is shown in formula (1): Among them, f a,b represents a sub-feature map, a∈(0,1,2…scale-1), b∈(0,1,2…scale-1), S represents the spatial size of the feature map, C1 represents the number of channels of the sub-feature map, and scale represents the number of sub-feature maps split out. The non-strided convolution layer is used to add a non-strided convolution layer using C2 convolution kernels to obtain the final feature map X" ( C2), where C2 <C1<scale 2 .

4. The method for identifying sick fry based on the improved YOLOv10 model according to claim 3, characterized in that: In step S2, the auxiliary branch includes a first auxiliary branch and a second auxiliary branch, the first auxiliary branch is connected in parallel with Stage1, and the second auxiliary branch is connected in parallel with Stage2; The first auxiliary branch includes an SPD-Conv module and an improved ShuffleBlock module connected in sequence, and the second auxiliary branch includes an improved ShuffleBlock module; the improved ShuffleBlock module includes an improved feature extraction unit and an improved downsampling unit; The improved feature extraction unit replaces the 1×1 convolution after the depth-separable convolution of the feature extraction unit in the basic module ShuffleBlock of ShuffleNetv2 with a partial convolution, wherein the partial convolution performs a standard convolution operation on part of the input channels to extract spatial features, and the remaining channels remain unchanged, and the partial ratio of the partial convolution is r=1 / 2; The improved downsampling unit is to replace the depth-separable convolution of the downsampling unit in the basic module ShuffleBlock of ShuffleNetv2 with an SPD-Conv module.

5. The method for identifying sick fry based on the improved YOLOv10 model according to claim 1, characterized in that: In step S2, the HFG-based fusion feature fusion method includes the following steps: Step a. Given a low-level feature map X low ∈R H×W×C and high-level feature map X high ∈R H×W×C , the high-level feature map X high ∈R H×W×C The channels are evenly split into G groups to obtain G groups of feature maps. If they cannot be evenly divided, they are rounded down. Step b. Perform 1×1Conv operation on each group of feature maps after splitting, so that the number of channels of each group of feature maps is the same as that of the low-level feature map X low The number of channels is the same, and each set of feature maps after dimensionality increase is recorded as X g ; Step c. Each set of feature maps X g Respectively with X low After adding each element, the feature graphs of each G group are accumulated to obtain the feature graph X′ high , as shown in formula (2): Among them, g∈(1,2…G); Step d. The feature map X′ high Global average pooling is performed in the spatial dimension to compress the entire spatial information into a global feature and obtain a C-dimensional global feature. The compression process is shown in formula (3): Among them, z c ∈R c represents the global feature, H represents the height of the feature map, W represents the width of the feature map, and u c represents the input matrix, i represents the row of the input matrix, j represents the column of the input matrix, and c represents the number of channels of the feature map; Step e. Pass the global feature through a fully connected layer consisting of a ReLU activation function and a Sigmoid activation function to obtain a weighted output, as shown in formula (4): s=σ(W2δ(W1z c )) (4) Where s represents the weight output, δ is the ReLU activation function, σ is the Sigmoid activation function, W1(·) represents the fully connected layer operation of the ReLU activation function, and W2(·) represents the fully connected layer operation of the Sigmoid activation function. Step f. Combine the weight output with the low-level feature map X low Perform element-by-element multiplication and concatenate with the high-level feature map along the channel to obtain the final output, as shown in formula (5): Among them, Out represents the final output, represents element-by-element multiplication; [·,·] represents a concatenation operation.

6. The method for identifying sick fry based on the improved YOLOv10 model according to claim 1, characterized in that: In step S2, the SlideLoss loss function is shown in formula (6): Among them, I n represents the loss of the nth category, n∈(1,2…n), f(x) represents the Slide weighting function added to the baseline model loss function, w n represents the weight factor, y n represents the nth actual value, σ represents the Sigmoid function, x n represents the nth predicted value, and x represents the function input; Based on the baseline model loss function, the Slide weighting function f(x) is added as shown in formula (7): Among them, μ represents the average value of the intersection-over-union ratio between all predicted boxes and the true boxes.

7. A method for identifying sick fry based on an improved YOLOv10 model according to any one of claims 1 to 6, characterized in that: In the step S3, the improved YOLOv10-UDFishNet model is trained based on the image data set until the training is completed when the set learning round is reached, thereby obtaining the trained YOLOv10-UDFishNet model.

8. The method for identifying sick fry based on the improved YOLOv10 model according to claim 1, characterized in that: The fish detection results obtained in step S4 include normal and abnormal.

Images