Underwater PLDNET network-based real-time Dongpo fish disease identification method

Abstract

The invention relates to an underwater PLDNET network-based real-time Dongpo fish disease identification method, which comprises the following steps: S1, data acquisition and preprocessing: acquiring a Dongpo fish disease image data set, and preprocessing the data; s2, feature extraction: inputting the preprocessed Dongquan star spot disease data into a feature extraction module, and obtaining a feature extraction result after multi-level deep convolution; s3, feature enhancement and fusion: inputting a feature-extracted Dongquan star spot disease data result into a feature enhancement and fusion module, and enhancing the expression of key information to obtain an enhancement and fusion result; s4, replacing a loss function, replacing an improved minimum point distance intersection-to-union ratio loss function, and accurately positioning a small-target east star spot disease area; and S5, predicting a disease result of the Dongguo, and outputting a plurality of disease prediction frames and categories of the Dongguo. According to the invention, the real-time detection of the disease of the Dongquasichthys nobilis is realized, the occurrence of the disease can be quickly responded, and the requirements of high-efficiency and accurate disease detection in aquaculture are met.

Description

Technical Field

[0001] The invention relates to the technical field of aquaculture disease identification, and more specifically to an underwater real-time grouper disease identification method based on a PLDNET network. Background Art

[0002] In the field of aquaculture, early monitoring and accurate diagnosis of fish diseases are crucial to ensuring the health of aquaculture and improving economic benefits. Take the grouper as an example. Due to its high market value and nutritional value, it is widely used in aquaculture. However, in a high-density equipment-based aquaculture environment, grouper is prone to outbreaks of diseases such as vibriosis and parasitic diseases, which poses a huge challenge to the aquaculture management of grouper. Once a disease breaks out, the high mortality rate of fish and the decline in the market value of the product will lead to serious economic losses.

[0003] In the prior art, fish disease detection methods are mainly divided into traditional manual naked eye detection and automated detection based on machine learning. The manual naked eye detection method relies on experts to make a diagnosis through naked eye observation. Although it has certain practical significance, it has the problems of strong subjectivity, long time consumption, low precision causing misjudgment, high labor cost, etc., which makes it difficult to meet the needs of large-scale breeding in the market. In recent years, with the continuous development of artificial intelligence, machine learning methods based on image processing technology have gradually been applied to fish disease detection. The automated detection method based on machine learning improves the efficiency and accuracy of fish disease detection through steps such as image segmentation, feature extraction and classification. However, the automated detection method based on machine learning still faces some limitations: First, when dealing with smaller lesion targets such as parasitic diseases and early vibriosis, the detection accuracy of existing machine learning methods is often insufficient, and it is difficult to accurately identify and locate the lesion area, resulting in limited early identification ability of fish diseases. Secondly, due to the lack of disease image data sets for specific fish, the generalization ability of existing models is limited and it is difficult to extend to different breeding environments. Finally, real-time performance is still an urgent problem to be solved. At present, most methods focus on improving the detection accuracy of models, which cannot meet the needs of rapid detection and real-time response in actual breeding environments. Therefore, in aquaculture of grouper, given its small disease target and complex background.

[0004] First, the existing automated detection technology based on machine learning does not perform well in complex aquaculture environments (such as changes in light and fluctuations in water quality), and it is difficult to accurately identify and locate smaller lesion areas, especially in the detection of micro-disease targets such as early vibriosis and parasitic diseases. This technical limitation makes it difficult to identify early diseases, often causing the disease to develop to the middle and late stages, ultimately causing greater economic losses to the breeding of grouper. Secondly, due to the lack of high-quality disease image datasets for specific fish (such as grouper), the existing target detection methods have significant limitations in generalization ability and cannot be effectively extended to specific grouper breeding environments. This limitation has led to a technical gap in the field of disease identification for grouper, and there is an urgent need to develop more adaptable detection solutions. Finally, although the current technology has improved detection accuracy in some cases, it performs poorly in real time and cannot respond to actual breeding needs in a timely manner, thus delaying disease management. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an underwater real-time grouper disease identification method based on PLDNET network, which is efficient, accurate and has real-time detection capabilities to meet the detection and positioning needs of early grouper diseases, while accurately identifying the types of diseases, thereby effectively responding to health management challenges in aquaculture.

[0006] The technical solution adopted by the present invention to solve the technical problem is to construct an underwater real-time grouper disease identification method based on PLDNET network, comprising the following steps:

[0007] S1. Data collection and preprocessing: collect disease image datasets of grouper fish and preprocess the data, including data enhancement, annotation and dataset division;

[0008] S2, feature extraction, input the pre-processed eastern star spot disease data into the feature extraction module in the PLDD network model, and obtain the feature extraction results after multi-level deep convolution;

[0009] S3, feature enhancement and fusion, input the data results of the eastern star spot disease after feature extraction into the feature enhancement fusion module, after the focus modulation layer and multi-scale feature fusion, the feature information of the eastern star spot disease from different levels is integrated, the expression of key information is enhanced, and the enhanced fusion result is obtained;

[0010] S4, replace the loss function, replace the traditional bounding box loss function with the improved minimum point distance intersection loss function, and accurately locate the small target oriental star spot disease area;

[0011] S5. Prediction of grouper disease results. The feature enhancement fusion results are input into the PLDNet feature prediction module for prediction, and multiple grouper disease prediction boxes and categories are output.

[0012] According to the above scheme, in step S1, data collection and preprocessing are implemented by the following method, specifically:

[0013] S101, using underwater cameras in grouper fisheries to collect grouper disease image datasets in real time;

[0014] S102, performing a variety of image enhancement operations such as mirroring, flipping, cropping, rotating, translating, blurring, adding noise and adjusting brightness on the images of the eastern grouper disease in the data set, and constructing a PLDD eastern grouper disease data set;

[0015] S103, using Labelme annotation software to accurately annotate the target disease areas in the enhanced dataset image one by one, including manually drawing the disease outline using key point annotations and assigning corresponding disease category labels to each annotated area to ensure that each disease area can be accurately labeled;

[0016] S104. Divide the PLDD dataset into a training set, a test set, and a validation set in a ratio of 7:2:1, wherein the training set is used for pre-training of the PLDD model, the test set is used for evaluating and adjusting the performance of the PLDD model, and the validation set is used to measure the final performance of the model and evaluate its generalization ability.

[0017] According to the above scheme, in step S2, feature extraction is achieved by the following method, specifically:

[0018] S201, inputting the preprocessed oriental spot disease image data into the PLDNet feature extraction network, wherein the feature extraction module is composed of a multi-layer convolutional layer and a residual block layer to extract different levels of features in the oriental spot image;

[0019] S202, Silence layer feature suppression: the oriental star spot disease image will first pass through the Silence layer when entering the feature extraction module, which is used to suppress certain noise or features unrelated to the disease in the input oriental star spot disease image, so as to improve the effect of feature extraction in subsequent layers;

[0020] S203, convolution layer feature extraction, after the Silence layer, multi-layer deep convolution operation is used to extract multiple local features of the disease edge, texture, and shape layer by layer, gradually reducing the image size of the eastern star spot disease, increasing the depth of the feature map, and capturing more abstract features;

[0021] S204, residual block layer feature optimization. In the feature extraction module, each convolution layer is followed by a residual block layer to improve the nonlinear representation ability of the model. The deep features of the eastern star spot disease image are further extracted through the residual block layer. The gradient vanishing problem is reduced by using jump connections. The perception of the eastern star spot disease area is improved through multi-scale extraction of local features.

[0022] S205: Output feature map. Finally, output a high-dimensional feature map after multi-layer convolution and residual block layer processing.

[0023] According to the above scheme, in step S3, feature enhancement and fusion are achieved by the following method, specifically:

[0024] S301, input the feature map output in step S2 into the feature fusion enhancement module to enhance the features of the eastern grouper disease feature map using the focus modulation layer mechanism, and the focus modulation layer is divided into three parts: Query, Modulator, Aggregate;

[0025] S302, feature fusion and upsampling, through multi-scale feature fusion and upsampling operations, feature maps at different levels are integrated and spatial resolution is gradually restored to ensure that the PLDNet network can process large, medium and small disease areas at the same time;

[0026] S303, outputting an enhanced fusion feature map after feature enhancement and feature fusion processing at the focal modulation layer, to provide an optimized input for subsequent prediction of the disease area of ​​grouper.

[0027] According to the above scheme, in step S301, in the feature fusion enhancement module, the output feature map of the eastern grouper disease is enhanced by using the focus modulation layer mechanism, specifically:

[0028] S301a, input the feature map to the Query part, and the Query part extracts the key area feature q(X) according to the input feature map X for subsequent focus modulation, which is expressed as:

[0029] q(X)=f q (X)

[0030] Among them, f q (·) is a mapping function used to obtain the key features of the input image;

[0031] S301b, aggregate and adaptively adjust the features at different levels through the Modulator part. First, extract the multi-level context information to form a feature set Z. The feature of each level is represented by Z l , which are aggregated and weighted by a linear transformation:

[0032] Zl =GeLU(Conv dw (Z l-1 ))

[0033] The final modulator output is obtained by weighted summation:

[0034]

[0035] Among them, G l is the weight, L is the number of feature layers, Indicates element-by-element multiplication, the final output Z out is the feature map after adaptive aggregation;

[0036] S301c, in the Aggregate part, the features output by the modulator are combined with the Query features through element-by-element affine transformation to further enhance the significance of the diseased area:

[0037]

[0038] Among them, g i l represents the weight, z i l is the feature of the lth layer, h(·) is the linear transformation function, y i The final enhancement feature.

[0039] According to the above scheme, in step S4, the method of replacing the traditional bounding box loss function with the improved minimum point distance intersection over union (MPDIoU) loss function is:

[0040] First, define the coordinates of the upper left corner and lower right corner of the real box and the predicted box, calculate the Euclidean distance between the corresponding vertices of the two bounding boxes, and at the same time, calculate the overlapping area and union area of ​​the predicted box and the real box, and standardize the distance between the vertices and divide it by the sum of the squares of the width and height of the bounding box. The calculation formula of MPDIoU is:

[0041]

[0042] Among them, the overlapping area of ​​the predicted box and the real box is A gt ∩A pre and the area of ​​the union A gt ∪A pre , define the real box A gt The coordinates of the upper left corner are (x 1,gt ,y 1,gt ), the coordinates of the lower right corner are (x 2,gt ,y 2,gt ); Similarly, predict box A pre The coordinates of the upper left corner are (x1,pre ,y 1,pre ), the coordinates of the lower right corner are (x 2,pre ,y 2,pre ), for the corresponding vertices of the two bounding boxes, calculate the Euclidean distance between the upper left corner and the lower right corner respectively:

[0043] d1 2 =(x 1,pre -x 1,gt ) 2 +(y 1,pre -y 1,gt ) 2

[0044] d2 2 =(x 2,pre -x 2,gt ) 2 +(y 2,pre -y 2,gt ) 2

[0045] w and h are the width and height of the predicted box or the true box respectively;

[0046] MPDIoU loss function L MDPIoU The goal is to minimize the distance between the predicted box and the real box while maximizing their overlapping area. The loss function expression is:

[0047] L MDPIoU =1-MDPIoU.

[0048] According to the above scheme, in step S5, the prediction of the disease results of grouper is achieved by the following method, specifically:

[0049] S501, inputting the feature map after feature enhancement fusion into the PLDNet feature prediction module to perform positioning and classification of the diseased area of ​​grouper;

[0050] S502, PLDNet feature prediction module generates three levels of prediction frames of different scales according to the input feature map, namely P3 / 8, P4 / 16, and P5 / 32. Each prediction frame is used to detect the diseased areas of small, medium, and large-sized grouper fish, ensuring that all types of diseased areas can be accurately located;

[0051] S503, for each prediction box, the model outputs the position coordinates through the box regression head, and assigns the grouper disease category label and the confidence interval of 0-1 to the prediction box through the cls classification head;

[0052] S504. The generated prediction frames are optimized and screened by a non-maximum suppression method to remove redundant prediction frames and prediction frames with low confidence, thereby obtaining a high-confidence detection result of the diseased area of ​​grouper.

[0053] The underwater real-time grouper disease identification method based on PLDNET network of the present invention has the following beneficial effects:

[0054] 1. The PLDNet (Plectropomus Leopardus Disease Detection Network) proposed in the present invention enhances the recognition accuracy of early small lesion areas by introducing a focal modulation layer, and improves the detection accuracy of lesion areas by optimizing the loss function through the improved minimum point distance intersection over union (MPDIoU). It provides a new PLDD (Plectropomus Leopardus Disease Dataset) data set, fills the gap in the image data of eastern star grouper disease, effectively improves the generalization ability of the model, and makes up for the technical gap in the field of eastern star grouper disease recognition based on deep learning. The PLDNet network architecture is proposed to realize real-time detection of eastern star grouper fish diseases, which can quickly respond to the occurrence of diseases and meet the needs of efficient and accurate disease detection in aquaculture;

[0055] 2. The present invention can improve the accuracy of identifying micro-disease areas of grouper. In view of the changes in illumination and water quality in complex aquaculture environments, PLDNet significantly improves the recognition accuracy of micro-lesion areas of grouper by introducing multi-scale feature fusion technology and improved loss function (MPDIoU). Experimental results show that PLDNet has an accuracy of 88.1% in the mAP@0.5 index, which is more than 8% higher than the existing advanced target detection models. In particular, PLDNet shows higher stability and reliability in the detection of small lesion targets such as early vibriosis and parasitic diseases, which significantly reduces the risk of early diseases being neglected and effectively reduces the economic losses of grouper caused by delayed diagnosis.

[0056] 3. The present invention improves real-time performance and reduces the computational complexity of the model by optimizing the PLDNet network structure and algorithm, thereby greatly improving the detection speed. In the experiment, PLDNet can achieve efficient real-time detection in actual application scenarios, which is about 15% faster than the response speed of the prior art. This rapid response capability not only meets the farmers' demand for real-time disease monitoring, but also improves the efficiency of grouper disease management and reduces further spread and losses caused by delayed detection.

[0057] 4. The present invention enhances the generalization ability by constructing a high-quality grouper disease image dataset (PLDD), which makes up for the deficiency of the existing technology in the scarcity of datasets. The introduction of the PLDD dataset enables the PLDNet model to show stronger generalization ability in different breeding environments, especially in the face of different lighting conditions and complex water quality backgrounds. PLDNet can maintain the stability of detection. Compared with the traditional model, the recognition accuracy of PLDNet in different environments is improved by about 10%, filling the technical gap of the current technology in the detection of grouper breeding diseases;

[0058] 5. The detection method provided by the present invention simplifies the operation process, is easy to operate, and is convenient to deploy, reducing the dependence on professional knowledge, so that farmers and technicians can easily apply it. Compared with the existing technology, PLDNet reduces the need for manual intervention through a highly automated real-time detection process, significantly improves the convenience of operation, and promotes the popularization and application of this technology in large-scale farms. This simplified process can help farmers make diagnostic decisions quickly, thereby improving the efficiency of the entire breeding management. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0060] Figure 1 It is a flow chart of the underwater real-time grouper disease identification method based on PLDNET network of the present invention;

[0061] Figure 2 This is an example diagram of the PLDD sample of the grouper fish disease dataset of the present invention;

[0062] Figure 3 This is an example diagram of enhanced data on grouper fish diseases according to the present invention;

[0063] Figure 4 It is an example diagram of the PLDNet network structure of the present invention;

[0064] Figure 5 1 is a diagram showing an example of the structure of a Focal Modulaiton module of the present invention;

[0065] Figure 6 This is an example diagram of the MPDIoU (intersection-over-union ratio based on minimum point distance) principle of the present invention;

[0066] Figure 7 This is an example diagram of grouper detection according to the present invention. DETAILED DESCRIPTION

[0067] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0068] like Figure 1-7 As shown, the underwater real-time grouper disease identification method based on PLDNET network of the present invention comprises the following steps:

[0069] S1. Data collection and preprocessing: Collect disease image datasets of grouper fish and preprocess the data, including data enhancement, labeling, and dataset division.

[0070] Taking Vibrio ostreatus infection as an example, the following steps are specifically included:

[0071] S101. Use underwater cameras in grouper fisheries to collect grouper disease image datasets in real time.

[0072] The images of Vibrio infection of grouper fish were taken in the natural breeding environment of the grouper fishery. In the breeding environment, natural light and artificial light were used to take pictures respectively to ensure that the disease images of grouper fish were obtained under different lighting conditions. At the same time, the diseased grouper fish were carefully observed and recorded, especially the specific parts and different infection stages of Vibrio infection. During the collection process, each grouper fish was photographed from multiple angles from the front, back and side to record the full picture of its Vibrio infection area. Figure 2 As shown in the figure, the grouper fish disease image dataset PLDD, the PLDD dataset has 1041 image data, which are mainly divided into three categories: normal grouper (a), Vibrio infection (b), and leech infection (c).

[0073] S102, perform various image enhancement operations such as mirroring, flipping, cropping, rotating, translating, blurring, adding noise, and adjusting brightness on the images of the eastern star spot disease in the data set. Three of the enhancement operations are randomly combined each time, and three images with different enhancement operations are generated for each original eastern star spot disease image to construct the PLDD eastern star spot disease data set.

[0074] Take an original image of a red spot infected with Vibrio as an example. First, flip it left and right, and then generate three different enhanced images of the red spot by rotating it 10 degrees, adding Gaussian noise, and adjusting the brightness. Each time, three of the above enhancement methods are randomly selected and randomly combined to expand the diversity of the data set and ensure that the model can maintain the ability to recognize Vibrio infection under different lighting, angles, and noise interference. Figure 3As shown in the figure, examples of data enhancement for grouper fish diseases, (a) is the original image of grouper, (b) is the data enhanced image of (a) generated after upside-down flipping, adding noise, and cropping, (c) is the data enhanced image of (a) generated after flipping, translating, and adjusting the brightness, (d) is the data enhanced image of (a) generated after adjusting the brightness, adding noise, and blurring.

[0075] S103, using Labelme annotation software to accurately annotate the target disease areas in the enhanced dataset image one by one. This step includes manually drawing the disease outline using key point annotations and assigning a corresponding disease category label to each annotated area to ensure that each disease area can be accurately labeled.

[0076] For each image of Vibrio infection, key point annotation is used to manually draw the outline of the Vibrio infection area to ensure that all Vibrio infection spots are accurately marked. Each disease area is also given a corresponding category label, such as "Vibrio infection disease". Through this precise annotation method, the model can learn the different disease characteristics of grouper fish during the training process, thereby improving the model's detection performance for various fish diseases.

[0077] S104, divide the PLDD dataset into a training set, a test set, and a validation set in a ratio of 7:2:1. The training set is used for pre-training of the PLDD model, the test set is used for evaluating the performance and adjusting the parameters of the PLDD model, and the validation set is used for measuring the final performance of the model and evaluating the generalization ability.

[0078] After the image annotation of the PLDD Vibrio sutchuenensis dataset is completed, the entire Vibrio sutchuenensis disease image dataset is divided into training set, test set and validation set in a ratio of 7:2:1. For example, assuming there are 1,000 annotated images of Vibrio sutchuenensis infection, 700 images are used as the training set for the PLDD model, 200 images are used as the test set to evaluate the preliminary performance of the PLDD model, and 100 images are used as the validation set to finally verify the generalization ability of the PLDD model. Through this reasonable data division, it is possible to ensure that the sample distribution of each dataset is consistent, avoid overfitting or underfitting of the model, and thus improve the prediction accuracy of the model for unknown data.

[0079] S2. Feature extraction: input the pre-processed Eastern Star Spot disease data into the feature extraction module in the PLDD network model, and obtain the feature extraction results after multi-level deep convolution. Specifically:

[0080] S201, input the preprocessed image data of the eastern star spot disease into the PLDNet feature extraction network. The feature extraction module is composed of multiple convolutional layers (Conv) and residual block layers (ResNCSPELAN4) to extract different levels of features in the eastern star spot image.

[0081] The image data of Vibrio infection in grouper fish after data enhancement and standardization is input into the feature extraction module. The PLDNet network consists of multiple convolutional layers (Conv) and residual block layers (ResNCSPELAN4). The PLDNet network structure is as follows: Figure 3 As shown in the figure. The convolutional layer first extracts low-level features from the image, such as identifying the edges, textures, and basic shapes of Vibrio fasciatus infection. Taking Vibrio fasciatus infection as an example, after the input image is processed by the initial convolutional layer, the model is able to identify the Vibrio infection area on the fish body. These low-level features provide a basis for subsequent deep analysis. The residual block layer allows the gradient to propagate deeper through jump connections, preventing the gradient from disappearing in the deep network. In addition, the residual block further extracts high-level features in the image, such as morphological changes and subtle pathological features of the diseased area. Taking leech infection as an example, the model can identify tiny lesions on the fish body through this network structure, especially in the early stage of symptom manifestations.

[0082] S202, Silence layer feature suppression. The oriental star spot disease image entering the feature extraction module will first pass through the Silence layer, which is used to suppress certain noise or disease-irrelevant features in the input oriental star spot disease image, aiming to improve the effect of feature extraction in subsequent layers.

[0083] In response to the infection of Vibrio sutchuenensis, the Silence layer first performs preliminary processing on the input image of the grouper fish, aiming to suppress the noise in the image and features unrelated to the disease of the grouper. By using a high-pass noise filtering algorithm, the Silence layer can identify and filter out elements such as light spots and suspended particles in the water that interfere with the identification of the grouper disease, so as to avoid affecting subsequent detection. In addition, the Silence layer can also weaken the features of areas in the image that are not related to the small target disease of the grouper (such as healthy grouper scale areas or uninfected parts of the fish body), thereby reducing the interference of non-lesion areas on detection. After these processes, the Silence layer will enhance the features related to the disease, especially the micro-lesion areas infected by Vibrio sutchuenensis, in order to improve the feature extraction accuracy of the subsequent convolutional layer. Finally, the Silence layer outputs the optimized feature map, providing a more accurate and effective basis for subsequent detection.

[0084] S203, convolutional layer feature extraction, after the Silence layer, multi-layer deep convolution (Conv) operation is used to extract various local features such as disease edge, texture, shape, etc. layer by layer, gradually reducing the image size of the eastern star spot disease and increasing the depth of the feature map, aiming to capture more abstract features.

[0085] In the process of convolutional layer feature extraction, multiple convolutional layers are used to gradually extract the features of the input image, and each convolutional layer captures more and more complex local features. The convolution kernel moves on the star spot image to detect edges, corners and local texture features.

[0086] Taking the Vibrio infection of the eastern grouper as an example, the convolution kernel slides on the image, and gradually mines the characteristics of the Vibrio infection in the image through layer-by-layer convolution operations. The initial convolution layer may only recognize the edge of the eastern grouper or the structure of the scales, which establishes a basic understanding of the geometric structure for the network. In the deeper convolution layers, the convolution kernel can focus on the tiny texture and color changes on the scales of the eastern grouper, and capture the local pathological characteristics related to Vibrio infection, such as white spots and ulcers caused by Vibrio infection.

[0087] When dealing with infections like Vibrio, the network not only needs to identify the area of ​​​​the disease, but also needs to remain robust in the face of complex backgrounds and changing lighting. Through the gradual extraction of layers of convolution, the network can identify subtle differences in the area of ​​​​the disease, and even the manifestation of lesions under different lighting conditions. Deeper convolutional layers can capture higher-dimensional features, such as the local abnormal diffusion caused by Vibrio infection, irregular changes in morphology, or the special white spot morphology on the scales of the eastern star spot. This multi-level feature extraction enables the model to still distinguish subtle lesions in the face of complex backgrounds, improving the detection accuracy and stability of the network in different environments.

[0088] S204, residual block layer feature optimization. In the feature extraction module, each convolutional layer is followed by a residual block layer to improve the nonlinear representation ability of the model. The deep features of the eastern star spot disease image are further extracted through the residual block layer (ResNCSPELAN4), and the gradient vanishing problem is reduced by using jump connections. The multi-scale extraction of local features aims to further improve the perception of the eastern star spot disease area.

[0089] The features extracted by the convolutional layer are optimized through the ResNCSPELAN4 residual block layer. The core function of the residual block layer is to capture the eastern star spot disease areas of different sizes and shapes through a multi-scale feature extraction mechanism. For example, the symptoms of leech infection are manifested as extremely small diseased areas, which may be difficult to identify using traditional convolutional layers. The residual block layer can ensure that the significant features of the leech disease area are fully learned through deeper feature extraction and jump connections, especially when the disease range is small. This optimization improves the model's ability to detect small target diseases. In addition, the residual block integrates information at different levels to ensure that each layer of convolutional features does not lose details, ultimately improving the stability and accuracy of the model.

[0090] S205, output feature map, finally, output high-dimensional feature map after multi-layer convolution and residual block layer processing. These feature maps have rich spatial and contextual information, aiming to provide a basis for subsequent feature enhancement and disease prediction.

[0091] After multi-layer convolution and residual block processing, the network outputs high-dimensional feature maps, which contain rich spatial details and contextual information in the image. For example, in the detection of infection of the eastern star spot leech, the output high-dimensional feature map can not only capture the main diseased area on the surface of the fish, but also capture the surrounding healthy tissue, providing rich input information for subsequent feature enhancement and disease location. These feature maps are fused through multi-layer information to ensure that the model can still accurately locate and identify the diseased area even under complex background or lighting conditions. Finally, these high-dimensional feature maps will be passed to the feature enhancement module to further improve the accuracy of disease detection.

[0092] S3, feature enhancement and fusion, input the data results of the eastern star spot disease after feature extraction into the feature enhancement fusion module, after the focal modulation layer (FocalModulation) and multi-scale feature fusion, the feature information of the eastern star spot disease from different levels is integrated, the expression of key information is enhanced, and the enhanced fusion result is obtained. Specifically:

[0093] S301. Input the feature map output in step S2 into the feature fusion enhancement module and use the focal modulation layer (Focal Modulation) mechanism to enhance the features of the eastern grouper disease feature map. The focal modulation layer is mainly divided into three parts: Query, Modulator, and Aggregate. The focal modulation layer adaptively adjusts the weight of the features according to the importance of the diseased area, suppresses background noise, enhances the significance of the diseased area, and focuses on specific eastern grouper image areas to process local details, especially those small target lesions that are easily overlooked.

[0094] In an image of a possible Vibrio infection on the scales of a grouper, the focus modulation layer will give a higher weight to the infected area and reduce the weight of healthy tissue or underwater background. For example, the early symptoms of leech infection appear as tiny leech areas on the fins. Through the focus modulation mechanism, these tiny lesions can be enhanced in the feature map, ensuring that the model can focus on these small-scale lesions even in a larger field of view. Figure 5 As shown, this is an example diagram of the FocalModulaiton (focus modulation) module structure.

[0095] In the feature fusion enhancement module, the focal modulation layer (Focal Modulation) mechanism is used to enhance the output feature map of the eastern grouper disease. First, the feature map is input into the Query part, which extracts the key area feature q(X) based on the input feature map X for subsequent focal modulation. The formula can be expressed as:

[0096] q(X)=f q (X)

[0097] Among them, f q (·) is a mapping function used to obtain the key features of the input image.

[0098] Next, the Modulator part aggregates and adaptively adjusts the features at different levels. First, extract multi-level context information to form a feature set Z. The features at each level are represented as Z l . They are aggregated and weighted by a linear transformation:

[0099] Z l =GeLU(Conv dw (Z l-1 ))

[0100] The final modulator output is obtained by weighted summation:

[0101]

[0102] Among them, G l is the weight, L is the number of feature layers, Represents element-by-element multiplication. The final output Z out It is the feature map after adaptive aggregation.

[0103] Finally, in the Aggregate part, the features output by the modulator are combined with the Query features through element-by-element affine transformation to further enhance the saliency of the diseased area:

[0104]

[0105] Among them, g i l represents the weight, z i l is the feature of the lth layer, h(·) is the linear transformation function, y i The final enhancement feature.

[0106] S302, feature fusion and upsampling, integrates feature maps at different levels through multi-scale feature fusion and upsampling operations (Upsample), and gradually restores spatial resolution. This is to ensure that the PLDNet network can simultaneously process large, medium, and small disease areas and improve multi-scale detection capabilities.

[0107] Taking the leech infection of the eastern star spot and the Vibrio infection as examples, the differences in lesion characteristics between the two are quite obvious. Leech infection usually manifests as small and scattered lesions with a small lesion range, which is difficult to be accurately captured by conventional feature extraction methods. However, the lesions of Vibrio infection usually cover a larger area, and their characteristics are large lesions with more spatial coherence. In the process of multi-scale feature fusion, the PLDNet model uses a multi-layer convolutional neural network to extract feature information of different scales, and combines the focus modulation mechanism to adaptively enhance the features and adjust the weights of important areas.

[0108] In the initial stage, the model extracts the detailed features of small lesions such as leech infection through deep convolution. In the initial feature map, the small lesions of leech infection may only be some local subtle differences. Through layer-by-layer convolution processing, these subtle lesions are gradually enhanced, thereby capturing the edges and local details of leech infection. For these small lesion areas, the model uses layer-by-layer feature extraction and downsampling operations to ensure that fine-grained features are not submerged in the convolutional neural network. At the same time, the focus modulation mechanism adaptively suppresses background noise, further improving the detection ability of small lesions.

[0109] When dealing with Vibrio infections, the area of ​​the lesions is large, so the PLDNet model needs to not only extract local details during feature extraction, but also restore higher spatial resolution through the upsampling process to ensure accurate capture of the overall lesion morphology. The upsampling operation combined with feature information of different scales can gradually restore large-area lesions of Vibrio infections, and organically combine low-level edge information and high-level contextual information through feature fusion to ensure the complete display of the lesion area. In this process, the focus modulation module can adaptively adjust the weight of the features according to the importance of the diseased area, enhance the lesion significance, and suppress redundant background, thereby effectively improving the detection effect of large-area lesions.

[0110] Through the fusion and enhancement of this multi-scale feature, the PLDNet model can accurately capture the disease characteristics of different sizes and shapes by adaptively adjusting the receptive field and feature fusion process, whether it is a localized subtle leech infection spot or a large-scale Vibrio infection lesion. This process not only improves the robustness of the model in multi-scale disease detection, but also enhances its generalization ability in complex environments, and can cope with the spatial scale changes of various lesion characteristics, ensuring the accuracy and reliability of the detection results.

[0111] S303, outputting an enhanced fusion feature map after feature enhancement and feature fusion processing at the focal modulation layer, aiming to provide an optimized input for subsequent prediction of the disease area of ​​grouper.

[0112] In an image of a grouper with multiple diseases, the enhanced feature map can clearly distinguish different disease areas, such as Vibrio infection on fish scales and leech infection lesions on fish fins. This not only allows the model to more accurately identify the type of each disease, but also provides more reliable boundary information for locating the diseased areas of the grouper. Through this process, the model can reduce the false alarm rate (such as misjudging healthy grouper areas as leech-infected areas) and improve the overall accuracy of disease detection.

[0113] S4. Replace the loss function and replace the traditional bounding box loss function with the improved minimum point distance intersection over union (MPDIoU) loss function to accurately locate the small target eastern star spot disease area.

[0114] Traditional IoU-like loss functions often fail to accurately locate grouper diseases when dealing with small target diseases such as leeches due to the small overlap between their bounding boxes, especially when the lesion area is complex in shape or inaccurate in position. MPDIoU introduces the concept of minimum point distance, which not only considers the overlapping area of ​​the bounding box, but also calculates the minimum distance between the corresponding points of the predicted box and the true box, thereby providing a more accurate metric for fine-grained feature positioning.

[0115] In the specific implementation, we first define the coordinates of the upper left corner and lower right corner of the real box and the predicted box, and calculate the Euclidean distance between the corresponding vertices of the two bounding boxes. At the same time, we calculate the overlapping area and union area of ​​the predicted box and the real box, and normalize the distance between the vertices by dividing it by the sum of the squares of the width and height of the bounding box. The calculation formula of MPDIoU is:

[0116]

[0117] Among them, the overlapping area of ​​the predicted box and the real box is A gt ∩A pre and the area of ​​the union A gt ∪A pre, define the real box A gt The coordinates of the upper left corner are (x 1,gt ,y 1,gt ), the coordinates of the lower right corner are (x 2,gt ,y 2,gt ); Similarly, predict box A pre The coordinates of the upper left corner are (x 1,pre ,y 1,pre ), the coordinates of the lower right corner are (x 2,pre ,y 2,pre ). For the corresponding vertices of the two bounding boxes, calculate the Euclidean distance between the upper left corner and the lower right corner respectively:

[0118] d1 2 =(x 1,pre -x 1,gt ) 2 +(y 1,pre -y 1,gt ) 2

[0119] d2 2 =(x 2,pre -x 2,gt ) 2 +(y 2,pre -y 2,gt ) 2

[0120] w and h are the width and height of the predicted box or the true box, respectively.

[0121] MPDIoU loss function L MDPIoU The goal is to minimize the distance between the predicted box and the true box while maximizing their overlapping area. The specific loss function expression is:

[0122] L MDPIoU =1-MDPIoU

[0123] By minimizing the MPDIoU loss function during training, the PLDNet model can more accurately locate small target disease areas of grouper, including those with more subtle lesion features, such as lesion areas of leech infection or Vibrio infection. This improvement not only optimizes the bounding box positioning of the disease area of ​​grouper, but also enhances the robustness of the PLDNet model in dealing with small target diseases and diseases in complex situations, and improves the detection performance of the PLDNet model.

[0124] S5. Prediction of grouper disease results: input the feature enhancement fusion results into the PLDNet feature prediction module for prediction, and output multiple grouper disease prediction boxes and categories. Specifically:

[0125] S501, input the feature map after feature enhancement fusion into the PLDNet feature prediction module to perform positioning and classification of the diseased area of ​​grouper.

[0126] When detecting leech infection in grouper, after focus modulation and feature fusion processing, the feature map has highlighted the diseased area and suppressed the background noise. At this time, the important disease information in the feature map is passed to the feature prediction module, which will further process and analyze these diseased areas and prepare to generate prediction boxes for the diseased areas and classify them.

[0127] S502, PLDNet feature prediction module generates three levels of prediction frames of different scales (P3 / 8, P4 / 16, P5 / 32) according to the input feature map. Each prediction frame is used to detect small, medium and large-sized disease areas of grouper, respectively, to ensure that all types of disease areas can be accurately located.

[0128] The PLDNet feature prediction network generates prediction boxes of different scales (P3 / 8, P4 / 16, P5 / 32) based on the input feature map, and each prediction box corresponds to a disease area of ​​​​the eastern star grouper of different sizes. P3 / 8 is used to detect smaller disease areas, such as the early lesions of white spot disease of the eastern star grouper, which are usually small in size and difficult to detect. P4 / 16 and P5 / 32 deal with diseases in medium and large areas, respectively, such as the widespread infection area of ​​vibriosis. Through this multi-scale prediction, the model ensures that both small-scale white spot disease and vibriosis covering a large area can be accurately detected and the corresponding prediction box can be generated.

[0129] S503, for each prediction box, the model outputs the position coordinates through box (regression head), and assigns the prediction box with a category label of grouper disease and a confidence level ranging from 0 to 1 through cls (classification head), in order to distinguish different grouper disease types and determine their confidence levels.

[0130] In order to improve the positioning accuracy of the predicted box, the model uses MPDIoU (intersection over union based on minimum point distance) as a metric for the similarity between the predicted box and the true box. Unlike the traditional IoU (intersection over union) method, MPDIoU calculates the minimum distance between the vertices of the predicted box and the true box, so that even small-scale disease areas (such as early Vibrio infection spots) can be accurately aligned. This method avoids the deviation of traditional IoU when dealing with small targets, especially in the detection of grouper diseases, and can significantly improve the accuracy of disease area positioning.

[0131] The role of MPDIoU is to optimize the regression loss function of the model, ensuring that the output prediction box not only has a high overlap with the real box, but also minimizes the shape and position differences between the two. For example, when detecting Vibrio infection, the traditional IoU may cause inaccurate positioning due to the irregular shape of the diseased area, while MPDIoU can accurately capture the boundaries of these irregular shapes, so that the prediction box is better aligned with the real diseased area. The principle of MPDIoU is as follows Figure 5 shown.

[0132] S504, optimizing and screening the generated prediction frames through the non-maximum suppression (NMS) method, removing redundant prediction frames and prediction frames with low confidence, so as to finally obtain a set of high-confidence grouper disease area detection results.

[0133] Some prediction boxes may have multiple overlapping boxes due to lighting, angle, etc. The NMS algorithm will retain the box with the highest confidence and remove other redundant boxes in the overlapping area to reduce the problem of multiple detections. In addition, prediction boxes with lower confidence (such as confidence less than 50%) will also be filtered out to ensure that the final output results contain a small number of high-confidence predictions of diseased areas. For example, after NMS processing, the leech-infected area on the fin of the eastern starfish may only have one high-precision prediction box left, and the confidence level reaches 95%, thus providing a reliable basis for subsequent decision-making, such as Figure 5 As shown, the specific prediction effect of grouper detection example diagram.

[0134] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. An underwater real-time grouper disease identification method based on PLDNET network, characterized in that: The following steps are involved: S1. Collect the disease image dataset of grouper fish and preprocess the data, including data enhancement, annotation and dataset division; S2, input the pre-processed eastern star spot disease data into the feature extraction module in the PLDD network model, and obtain the feature extraction results after multi-level deep convolution; S3. The data results of the eastern star leaf spot disease after feature extraction are input into the feature enhancement fusion module. After the focus modulation layer and multi-scale feature fusion, the feature information of the eastern star leaf spot disease from different levels is integrated to enhance the expression of key information and obtain the enhanced fusion result. S4, replace the traditional bounding box loss function with the improved minimum point distance intersection loss function to accurately locate the small target oriental star spot disease area; S5. Prediction of grouper disease results. The feature enhancement fusion results are input into the PLDNet feature prediction module for prediction, and multiple grouper disease prediction boxes and categories are output.

2. The underwater real-time grouper disease identification method based on PLDNET network according to claim 1 is characterized in that: In step S1, data collection and preprocessing are implemented by the following method, specifically: S101, using underwater cameras in grouper fisheries to collect grouper disease image datasets in real time; S102, performing a variety of image enhancement operations such as mirroring, flipping, cropping, rotating, translating, blurring, adding noise and adjusting brightness on the images of the eastern grouper disease in the data set, and constructing a PLDD eastern grouper disease data set; S103, using Labelme annotation software to accurately annotate the target disease areas in the enhanced dataset image one by one, including manually drawing the disease outline using key point annotations and assigning corresponding disease category labels to each annotated area to ensure that each disease area can be accurately labeled; S104. Divide the PLDD dataset into a training set, a test set, and a validation set in a ratio of 7:2:1, wherein the training set is used for pre-training of the PLDD model, the test set is used for evaluating and adjusting the performance of the PLDD model, and the validation set is used to measure the final performance of the model and evaluate its generalization ability.

3. The underwater real-time grouper disease identification method based on PLDNET network according to claim 1 is characterized in that: In step S2, feature extraction is achieved by the following method, specifically: S201, inputting the preprocessed oriental spot disease image data into the PLDNet feature extraction network, wherein the feature extraction module is composed of a multi-layer convolutional layer and a residual block layer to extract different levels of features in the oriental spot image; S202, Silence layer feature suppression: the oriental star spot disease image will first pass through the Silence layer when entering the feature extraction module, which is used to suppress certain noise or features unrelated to the disease in the input oriental star spot disease image, so as to improve the effect of feature extraction in subsequent layers; S203, convolution layer feature extraction, after the Silence layer, multi-layer deep convolution operation is used to extract multiple local features of the disease edge, texture, and shape layer by layer, gradually reducing the image size of the eastern star spot disease, increasing the depth of the feature map, and capturing more abstract features; S204, residual block layer feature optimization. In the feature extraction module, each convolution layer is followed by a residual block layer to improve the nonlinear representation ability of the model. The deep features of the eastern star spot disease image are further extracted through the residual block layer. The gradient vanishing problem is reduced by using jump connections. The perception of the eastern star spot disease area is improved through multi-scale extraction of local features. S205: Output feature map. Finally, output a high-dimensional feature map after multi-layer convolution and residual block layer processing.

4. The underwater real-time grouper disease identification method based on PLDNET network according to claim 1 is characterized in that: In step S3, feature enhancement and fusion are achieved by the following method, specifically: S301, input the feature map output in step S2 into the feature fusion enhancement module to enhance the features of the eastern grouper disease feature map using the focus modulation layer mechanism, and the focus modulation layer is divided into three parts: Query, Modulator, Aggregate; S302, feature fusion and upsampling, through multi-scale feature fusion and upsampling operations, feature maps at different levels are integrated and spatial resolution is gradually restored to ensure that the PLDNet network can process large, medium and small disease areas at the same time; S303, outputting an enhanced fusion feature map after feature enhancement and feature fusion processing at the focal modulation layer, to provide an optimized input for subsequent prediction of the disease area of ​​grouper.

5. The underwater real-time grouper disease identification method based on PLDNET network according to claim 4 is characterized in that: In step S301, in the feature fusion enhancement module, the output feature map of the eastern grouper disease is enhanced by using the focus modulation layer mechanism, specifically: S301a, input the feature map to the Query part, and the Query part extracts the key area feature q(X) according to the input feature map X for subsequent focus modulation, which is expressed as: q(X)=f q (X) Among them, f q (·) is a mapping function used to obtain the key features of the input image; S301b, aggregate and adaptively adjust the features at different levels through the Modulator part. First, extract the multi-level context information to form a feature set Z. The feature of each level is represented by Z l , which are aggregated and weighted by a linear transformation: WITH l =GeLU(Conv dw (WITH l-1 )) The final modulator output is obtained by weighted summation: Among them, G l is the weight, L is the number of feature layers, Indicates element-by-element multiplication, the final output Z out is the feature map after adaptive aggregation; S301c, in the Aggregate part, the features output by the modulator are combined with the Query features through element-by-element affine transformation to further enhance the significance of the diseased area: Among them, g i l represents the weight, z i l is the feature of the lth layer, h(·) is the linear transformation function, y i The final enhancement feature.

6. The underwater real-time grouper disease identification method based on PLDNET network according to claim 1 is characterized in that: In step S4, the method of replacing the traditional bounding box loss function with the improved minimum point distance intersection-over-union (MPDIoU) loss function is as follows: First, define the coordinates of the upper left corner and lower right corner of the real box and the predicted box, calculate the Euclidean distance between the corresponding vertices of the two bounding boxes, and at the same time, calculate the overlapping area and union area of ​​the predicted box and the real box, and standardize the distance between the vertices and divide it by the sum of the squares of the width and height of the bounding box. The calculation formula of MPDIoU is: Among them, the overlapping area of ​​the predicted box and the real box is A gt ∩A pre and the area of ​​the union A gt ∪A pre , define the real box A gt The coordinates of the upper left corner are (x 1,gt ,y 1,gt ), the coordinates of the lower right corner are (x 2,gt ,y 2,gt ); Similarly, predict box A pre The coordinates of the upper left corner are (x 1,pre ,y 1,pre ), the coordinates of the lower right corner are (x 2,pre ,y 2,pre ), for the corresponding vertices of the two bounding boxes, calculate the Euclidean distance between the upper left corner and the lower right corner respectively: d1 2 =(x 1,pre -x 1,gt ) 2 +(y 1,pre -y 1,gt ) 2 d2 2 =(x 2,pre -x 2,gt ) 2 +(y 2,pre -y 2,gt ) 2 w and h are the width and height of the predicted box or the true box respectively; MPDIoU loss function L MDPIoU The goal is to minimize the distance between the predicted box and the real box while maximizing their overlapping area. The loss function expression is: L MDPIoU =1-MDPIoU.

7. The underwater real-time grouper disease identification method based on PLDNET network according to claim 1 is characterized in that: In step S5, the prediction of the disease results of grouper is achieved by the following method, specifically: S501, inputting the feature map after feature enhancement fusion into the PLDNet feature prediction module to perform positioning and classification of the diseased area of ​​grouper; S502, PLDNet feature prediction module generates three levels of prediction frames of different scales according to the input feature map, namely P3 / 8, P4 / 16, and P5 / 32. Each prediction frame is used to detect the diseased areas of small, medium, and large-sized grouper fish, ensuring that all types of diseased areas can be accurately located; S503, for each prediction box, the model outputs the position coordinates through the box regression head, and assigns the grouper disease category label and the confidence interval of 0-1 to the prediction box through the cls classification head; S504. The generated prediction frames are optimized and screened by a non-maximum suppression method to remove redundant prediction frames and prediction frames with low confidence, thereby obtaining a high-confidence detection result of the diseased area of ​​grouper.

Images