Novel fish acute septicemia detection system based on hyperspectral imaging

Abstract

The embodiment of the invention discloses a novel fish acute septicemia detection system based on hyperspectral imaging. The system comprises a data acquisition module, a model detection module and a user interaction module, wherein the data acquisition module acquires hyperspectral image data on the surface of a target fish body; the model detection module carries out diagnosis by taking the hyperspectral image data as input based on a trained deep learning diagnosis model, and the deep learning diagnosis model obtains a pixel space structure and a pixel spectrum sequence of the fish body surface based on the hyperspectral image data; identifying spatial features representing related clinical symptoms based on a pixel spatial structure, and identifying spectral features representing an inositol metabolism condition based on a pixel spectral sequence; and the user interaction module outputs a diagnosis result of the target fish body based on a result output by the model. The characteristic that the aeromonas hydrophila utilizes inositol as a unique carbon source and the map change of typical infection symptoms on the surface of a fish body are combined, and accurate and rapid detection is achieved through hyperspectral imaging and a deep learning algorithm.

Description

Technical Field

[0001] Multiple embodiments of this specification relate to the technical field of aquatic disease detection, specifically to the optimization of the convenience and efficiency of detecting novel acute septicemia in fish. Background Art

[0002] The channel catfish (Ictalurus punctatus), also known as the channel catfish and the bullhead catfish, belongs to the family Ictaluridae of the order Siluriformes. It is native to North America and is a large freshwater fish with excellent growth performance, strong stress resistance, and high-quality meat. Since it was introduced into China in 1984, it has been widely cultured across the country and is one of the important aquaculture economic fish species in China. With the increase in intensive aquaculture, the degradation of germplasm resources, and the abuse of antibiotics, the outbreaks of channel catfish diseases have shown an increasing trend year by year. Aeromonas hydrophila, belonging to the family Vibrionaceae and the genus Aeromonas, is a typical pathogen of zoonosis, and most fish infected with Aeromonas hydrophila show typical septicemia. In recent years, the novel acute septicemia of channel catfish caused by virulent Aeromonas hydrophila (vAh) has frequently broken out in various places, causing huge losses to the channel catfish aquaculture industry.

[0003] Traditional detection methods for channel catfish septicemia rely on laboratory physiological and biochemical and molecular biological detection methods, such as PCR detection (Polymerase Chain Reaction). By sampling through invasive fish body, then extracting genetic material for amplification, and using a probe with a fluorescent label for intensity detection. When the fluorescence intensity exceeds the threshold, it is like an alarm going off, confirming the detection of the virus. It has the advantages of high sensitivity and strong specificity.

[0004] However, the related methods such as PCR detection have the following defects: 1. The detection operation is cumbersome and requires high knowledge reserves and technical levels of the operators; 2. The cost of single detection is high; 3. It takes a long time. Usually, it takes 2 - 6 hours from sampling to obtaining the detection result, showing a certain lag, which is very inconvenient for large-scale fish body screening scenarios; 4. It requires destructive sampling. Thus, it cannot meet the requirements of modern aquaculture for instant, accurate, and efficient screening of aquatic diseases. Summary of the Invention

[0005] The embodiments of this specification provide a detection system for novel acute septicemia in fish based on hyperspectral imaging. Combining the characteristics of Aeromonas hydrophila using inositol as the sole carbon source and the spectral changes of typical infection symptoms on the fish body surface, through hyperspectral imaging and deep learning algorithms, rapid, non-destructive, and real-time detection of fish bodies can be achieved under the premise of high detection accuracy, thereby avoiding the spread of diseases.

[0006] The technical solution is as follows: In a first aspect, an embodiment of the present specification provides a novel acute septicemia detection system for fish based on hyperspectral imaging, including a data acquisition module, a model detection module, and a user interaction module: The data acquisition module acquires hyperspectral image data of the surface of the target fish body; The model detection module uses the trained deep learning diagnostic model to perform diagnosis with the hyperspectral image data as input, including an input layer, a convolutional layer, a feature fusion layer, a first fully connected layer, and a second fully connected layer; The input layer obtains the hyperspectral image data of the surface of the target fish body; The convolutional layer obtains the pixel spatial structure and pixel spectral sequence on the surface of the fish body based on the hyperspectral image data, and identifies spatial features representing relevant clinical symptoms based on the pixel spatial structure and spectral features representing inositol metabolism based on the pixel spectral sequence; The feature fusion layer performs feature fusion based on the spatial features and spectral features to obtain a fused feature map; The first fully connected layer compresses the dimension of the fused feature map to obtain a multi-dimensional vector corresponding to the number of feature variables, and obtains an adjusted multi-dimensional vector corresponding to the multi-dimensional vector according to the ReLU activation function; The second fully connected layer obtains a detection score based on the adjusted multi-dimensional vector and the weight matrix, and obtains a disease probability value corresponding to the detection score according to the Sigmoid activation function; The user interaction module outputs a diagnosis result of the target fish body based on the disease probability value and a first preset threshold.

[0007] By adopting the above technical solution, the data acquisition module acquires the hyperspectral image data of the target fish body and inputs it into the trained deep learning diagnostic model for feature extraction and recognition, and judges whether the target fish body is infected according to the predicted result output by the model. Hyperspectral imaging technology can capture the reflection characteristics of an object at different wavelengths, thereby providing detailed spectral and image information about the chemical properties of tissues or surfaces. Utilizing the metabolic characteristics that highly pathogenic Aeromonas hydrophila can utilize inositol as the sole carbon source, and the common clinical symptoms in fish after being infected with Aeromonas hydrophila, including petechial hemorrhage and fading starting from the tail, combined with deep learning algorithms to achieve accurate, rapid, and non-destructive identification and detection.

[0008] The detection only requires image acquisition of the fish body surface, without invasive sampling operations such as penetrating the fish body. The feature extraction and recognition speed of the model is fast and efficient. Although it costs to train the model, the cost of using the trained model for detection is very low, the operation is simple, the technical level requirements for operators are low, and the detection accuracy of the model will also increase with the continuous increase of the number of detections. The fusion and recognition of spatial features and spectral features ensure the accuracy of model detection.

[0009] As a preferred solution, the data acquisition module includes an acquisition unit and a control unit; The acquisition unit includes a black box, a halogen light source disposed inside the black box, a conveyor belt, a CCD camera located above the conveyor belt, and a spectrometer connected to the CCD camera; The control unit is connected to the spectrometer.

[0010] As a preferred solution, the convolutional layer includes a data separation unit, a spatial feature extraction unit, and a spectral feature extraction unit; The data separation unit obtains the pixel spatial structure and pixel spectral sequence on the fish body surface based on the hyperspectral image data; The spatial feature extraction unit uses the expert annotation method to obtain clinical symptom recognition labels including punctate hemorrhage, tail fading, and rot, and recognizes the spatial features in the pixel spatial structure based on the clinical symptom recognition labels; The spectral feature extraction unit uses the competitive adaptive reweighted sampling method, the successive projections algorithm, and the random frog leaping method to respectively identify the spectral feature wavelengths related to the inositol metabolism condition in the pixel spectral sequence from a preset wavelength range to obtain spectral features.

[0011] As a preferred solution, the feature fusion layer performs feature fusion based on the spatial feature, the spectral feature, and the dynamic weighting coefficient to obtain a fused feature map, and the dynamic weighting coefficient is adjusted based on the detection accuracy of the model detection module.

[0012] As a preferred solution, the adjusted multi-dimensional vector includes a plurality of one-dimensional vectors corresponding one-to-one to the feature variables; The second fully connected layer includes a mode selection unit, a first calculation unit, a second calculation unit, and a function unit; The mode selection unit selects the calculation mode of the detection score based on the compliance ratio of the one-dimensional vectors with qualified detection accuracy among all the one-dimensional vectors to all the one-dimensional vectors. Among them, the detection accuracy of each one-dimensional vector is updated in real time. When the compliance ratio is greater than the second preset threshold, the calculation mode is the preferred mode. When the compliance ratio is less than or equal to the second preset threshold, the calculation mode is the global mode; The first calculation unit obtains a first detection score based on the one-dimensional vectors with qualified detection accuracy among all the one-dimensional vectors and the weight values in their respective corresponding weight matrices in the preferred mode; The second calculation unit obtains a second detection score based on all the one-dimensional vectors and the weight matrix in the global mode; The function unit obtains the disease probability value of the first detection score or the second detection score according to the Sigmoid activation function.

[0013] As a preferred solution, the user interaction module also outputs the coordinates of the diseased area of the target fish body based on spatial features and spectral features.

[0014] As a preferred solution, the model detection module further includes a model training unit; The model training unit obtains a training data set and trains a model to be trained including a one-dimensional depth convolutional neural network and an image convolutional neural network based on the training data set to obtain a trained deep learning diagnosis model; Among them, the training data set includes multiple training samples corresponding to fish bodies with different disease levels. Each training sample includes hyperspectral image data on the surface of the fish body and the corresponding infection situation of the fish body, and the infection situation is determined by PCR detection.

[0015] As a preferred solution, obtaining the training data set in the model training unit includes: Obtain the average value of the data collected three times by the data acquisition module under constant temperature and humidity conditions as the hyperspectral image data in the training sample.

[0016] As a preferred solution, obtaining the training data set in the model training unit further includes: Correct the average value of the data collected three times by the data acquisition module under constant temperature and humidity conditions based on the dark reference image and the white reference image to obtain hyperspectral image data; Among them, the dark reference image is the image data collected by the CCD camera when the reflectivity is 0, and the white reference image is the image data collected by the CCD camera when the reflectivity is 100.

[0017] As a preferred solution, the model training unit uses the cross-entropy loss function to judge the training situation of the model to be trained.

[0018] The beneficial effects brought by the technical solutions provided in some embodiments of this specification at least include: 1. Obtain detailed hyperspectral image data on the surface of the fish body through the hyperspectral imaging system, and analyze the spectral and image information of the data through the deep learning diagnosis model to identify the features related to Aeromonas hydrophila infection, so as to be able to provide high-precision pathogen detection in real time without causing damage to fish, greatly improving the disease monitoring and management efficiency of aquaculture farms, accelerating the improvement of the construction of the aquaculture animal health management system, and ensuring the production and safety of aquatic products.

[0019] 2. The dual-modal feature fusion mechanism combines the inositol metabolism spectral fingerprint of the vAh strain with the image features of the diseased area for the first time to judge whether the target fish body is infected with Aeromonas hydrophila, improving the sensitivity of early detection.

[0020] 3. The model outputs the detection results in a way of global-submodel collaboration. The submodel focuses on specific features (spatial or spectral), while the global model makes comprehensive decisions to reduce the error of a single model and generally reduce the false alarm rate of the model. Description of the Drawings

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0022] Figure 1 It is a schematic structural diagram of a new type of acute septicemia detection system for fish based on hyperspectral imaging provided by the embodiments of this specification.

[0023] Figure 2 It is a schematic diagram of the deep learning diagnosis model architecture of the model detection module in a new type of acute septicemia detection system for fish based on hyperspectral imaging provided by the embodiments of this specification.

[0024] Figure 3 It is a schematic structural diagram of the data acquisition module in a new type of acute septicemia detection system for fish based on hyperspectral imaging provided by the embodiments of this specification.

[0025] In the figure: 1. Detection system; 100. Data acquisition module; 110. Acquisition unit; 111. Black box; 112. Halogen light source; 113. Conveyor belt; 114. Spectrometer; 115. CCD camera; 120. Control unit; 200. Model detection module; 210. Input layer; 220. Convolution layer; 230. Feature fusion layer; 240. First fully connected layer; 250. Second fully connected layer; 300. User interaction module; 400. Fish body. Detailed Embodiments

[0026] The following will clearly and completely describe the technical solutions in the embodiments of this specification in combination with the drawings in the embodiments of this specification.

[0027] The terms "first", "second", "third", etc. in the specification, claims and above drawings of this specification are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally further includes steps or units not listed, or optionally further includes other steps or units inherent to these processes, methods, products or devices.

[0028] The following description provides examples and does not limit the scope, applicability, or examples set forth in the claims. Changes may be made to the function and arrangement of the elements described without departing from the scope of the content of this specification. Various processes or components may be appropriately omitted, substituted, or added to each example. For example, the methods described may be performed in a different order than described, and various steps may be added, omitted, or combined. In addition, the features described for some examples may be combined into other examples.

[0029] A novel acute septicemia detection system for fish based on hyperspectral imaging, referring to Figure 1 、 Figure 2 as shown, Figure 1 shows a schematic structural diagram of the detection system 1, Figure 2 shows a schematic diagram of the deep learning diagnostic model architecture of the model detection module 200.

[0030] The detection system 1 includes a data acquisition module 100, a model detection module 200, and a user interaction module 300: The data acquisition module 100 acquires hyperspectral image data on the surface of the target fish body 400; The model detection module 200 performs diagnosis based on the trained deep learning diagnostic model with the hyperspectral image data as input, including an input layer 210, a convolutional layer 220, a feature fusion layer 230, a first fully connected layer 240, and a second fully connected layer 250; The input layer 210 acquires hyperspectral image data on the surface of the target fish body 400; The convolutional layer 220 acquires the pixel spatial structure and pixel spectral sequence on the surface of the fish body 400 based on the hyperspectral image data, and identifies the spatial features characterizing relevant clinical symptoms based on the pixel spatial structure and the spectral features characterizing the inositol metabolism situation based on the pixel spectral sequence; The feature fusion layer 230 performs feature fusion based on the spatial features and spectral features to obtain a fused feature map; The first fully connected layer 240 compresses the dimension of the fused feature map to obtain a multi-dimensional vector corresponding to the number of feature variables, and obtains an adjusted multi-dimensional vector corresponding to the multi-dimensional vector according to the ReLU activation function; The second fully connected layer 250 obtains a detection score based on the adjusted multi-dimensional vector and the weight matrix, and obtains a disease probability value corresponding to the detection score according to the Sigmoid activation function; The user interaction module 300 outputs a diagnosis result of the target fish body 400 based on the disease probability value and a first preset threshold.

[0031] Explanatory, myo-inositol, as a key metabolic intermediate and signaling molecule, plays an important role in the cell metabolism of Aeromonas hydrophila, helping it gain an ecological advantage in the complex aquatic environment. However, the metabolic characteristics and early symptoms of the highly toxic vAh strain are difficult to capture by conventional means. Different from traditional Aeromonas hydrophila, the characteristic that highly pathogenic Aeromonas hydrophila can utilize inositol as the sole carbon source provides a new breakthrough for early detection. Different from the traditional PCR detection method, in this embodiment, hyperspectral imaging is used to capture the reflection characteristics of the target fish body 400 at different wavelengths, so as to provide detailed spectral and image information about the tissue or surface chemical properties, thereby monitoring the spectral changes related to the metabolic process of the surface of the fish body 400 and the infected strain. Combining the image features of typical infection symptoms on the surface of the fish body 400, deep learning technology is used to automatically identify the infection features in the spectral and image data, providing rapid and accurate pathogen detection, and also enabling the system to detect infections at an early stage, thus avoiding the spread of diseases.

[0032] It should be noted that the data acquisition module 100 acquires the hyperspectral image data of the target fish body 400 through the spectrometer 114. Among them, each hyperspectral image is a three-dimensional data cube with dimensions of H*W*λ, where H and W are the height and width of the image, and λ is the wavelength. Prior to this, the target fish body 400 needs to be simply statically placed and dried to remove the excess moisture on the surface of the fish body 400 to ensure that the sample meets the requirements of spectral imaging. The convolutional layer 220 of the deep learning diagnostic model first extracts the pixel spatial structure and pixel spectral sequence on the surface of the fish body 400 from the hyperspectral image data, and then respectively performs feature recognition of relevant clinical symptoms and inositol metabolism conditions. The feature fusion layer 230 fuses the two types of features, and the fully connected layer is used to map the high-dimensional feature vector to the classification space. Nonlinearity is introduced to enhance the expression ability of the model. Among them, the ReLU activation function (f(x)=max(0,x)) can help alleviate the problem of gradient disappearance, accelerate the training process, and the output range is [0, +∞). After being activated by ReLU, non-negative and sparse high-level features are obtained. While using the Sigmoid activation function, the output value is between 0 and 1 and is used for binary classification (infected or not), converting the linear output into a probability value for easy judgment of the category. Finally, the user interaction module 300 compares the disease probability value output by the model with the first preset threshold (such as 0.5). When the disease probability value is greater than or equal to the first preset threshold, the diagnostic result of the target fish body 400 being infected is output, and when the disease probability value is less than the first preset threshold, the diagnostic result of the target fish body 400 not being infected is output.

[0033] In one embodiment of the present specification, the detection system 1 further includes a data input module connected to the model detection module 200, which receives hyperspectral image data uploaded to the server platform of the detection system 1 through a terminal to implement an online diagnosis function.

[0034] Explanatory, when performing the health detection of channel catfish, the user only needs to upload the hyperspectral image of the channel catfish to the data input module. After being calculated by the deep learning diagnosis model, the diagnosis result of the model can be obtained in real time, and the diagnosis result of whether the fish body 400 is healthy or infected with Aeromonas hydrophila can be output.

[0035] In some embodiments of the present specification, referring to the attached Figure 3 , the data acquisition module 100 includes an acquisition unit 110 and a control unit 120; The acquisition unit 110 includes a black box 111, a halogen light source 112 disposed inside the black box 111, a conveyor belt 113, a CCD camera 115 located above the conveyor belt 113, and a spectrometer 114 connected to the CCD camera 115; The control unit 120 is connected to the spectrometer 114.

[0036] Explanatory, the black box 111 is used to provide a pure spectral image data acquisition environment to eliminate external interference. The halogen light source 112 improves the stable illumination inside the black box 111. The conveyor belt 113 carries the target fish body 400 to move to the specified acquisition position. The control unit 120, such as a general control computer, serves as a coordination device for the data acquisition module 100 to control the spectrometer 114 to acquire the spectral image data of the target fish body 400.

[0037] In some embodiments of the present specification, the convolutional layer 220 includes a data separation unit, a spatial feature extraction unit, and a spectral feature extraction unit; The data separation unit obtains the pixel spatial structure and pixel spectral sequence on the surface of the fish body 400 based on the hyperspectral image data; The spatial feature extraction unit uses the expert annotation method to obtain clinical symptom recognition labels including punctate hemorrhage, tail fading, and rot, and recognizes the spatial features in the pixel spatial structure based on the clinical symptom recognition labels; The spectral feature extraction unit uses the competitive adaptive reweighted sampling method, the successive projection algorithm, and the random frog leaping method to respectively identify the spectral feature wavelengths related to inositol metabolism in the pixel spectral sequence within a preset wavelength range to obtain spectral features.

[0038] Explanatory, the pixel spatial structure describes the spatial structure of each pixel collected, and the pixel spectral sequence describes the spectral sequence of each pixel collected. The convolutional layer 220 adopts an attention mechanism: the attention mechanism of the spectral feature branch. Considering the metabolic characteristic that Aeromonas hydrophila utilizes inositol as the sole carbon source, selective analysis is carried out through specific bands related to the inositol metabolic process in the hyperspectral image. The attention mechanism of the spatial feature branch mainly focuses on the local lesion areas on the surface of the symptomatic diseased fish body 400, such as petechial hemorrhage, discolored areas, etc. These areas may be early signals of Aeromonas hydrophila infection. For these fine-grained features, the local areas of the infection symptoms can be focused on.

[0039] Illustrative, after Aeromonas hydrophila infects the fish body 400, common clinical symptoms include petechial hemorrhage, fading starting from the tail, etc. These pathological features are usually accompanied by immune responses and tissue damage, resulting in changes in the spectral characteristics on the surface of the fish body 400. For example, petechial hemorrhage may cause changes in the blood components in the tissue, thus generating different absorption peaks in the hyperspectral image. The fading symptom reflects pigment changes, which may lead to changes in reflectance and image features in the visible or near-infrared spectral range. The manifestation of these symptoms provides rich feature data for the hyperspectral imaging system, further improving the detection accuracy. Among them, the expert annotation method is to add labels to the data to provide supervision signals (such as labels, regional masks), clearly guiding the key features or categories that the model should focus on, thereby improving the recognition performance of the model. In addition, the data quality can be optimized by preprocessing the data, enhancing the contrast to make some features more obvious, or increasing data diversity through data augmentation. The three methods together improve the recognition performance of the model.

[0040] Illustrative, the spectral feature wavelengths related to the inositol metabolism situation are 400 - 1000 nm. The characteristic wavelengths obtained by CARS (Competitive Adaptive Reweighted Sampling), SPA (Successive Projection Algorithm), and RF (Random Frog Leap Method) are shown in Table 1. The features of these bands have unique significance for the judgment of early infection. When optimizing the input data of the model, these key bands are preferably selected to enhance the sensitivity of the model to the early stage of infection. Number of features Characteristic wavelength SPA 10 415,510,573,628,700,750,806,841,907,981 CARS 14 413,448,501,600,640,708,714,725,728,744,750,822,850,886 RF 11 435,501,568,608,643,776,780,900,912,928,937

[0041] Table 1: Characteristic wavelengths related to inositol of chondrichthyes obtained by different feature selection models.

[0042] Exemplarily, the spatial feature extraction branch includes a first spatial convolutional layer, a first spatial pooling layer, a second spatial convolutional layer, and a second spatial pooling layer. The input of the first spatial convolutional layer is the pixel spatial structure in the hyperspectral image data, and the spatial attention mechanism is the multi-dimensional tensor spatial features extracted after the image is convolved. Given an input feature vector, the model calculates three vectors through a linear transformation: Query (Q), Key (K), and Value. The calculation of each vector can be performed through different weight matrices. Then, the attention weights are calculated: the attention score is obtained by calculating the similarity between Query and Key. The formula is as follows: where Q is Query, K is Key, and dk is the dimension of Key. The weighted sum of all Value vectors is calculated, and the weights are determined by the attention scores calculated above. The final output is a weighted average representation. Finally, the Softmax function is used to normalize the weights so that the sum of the weights is 1. The connection relationship of the attention mechanism in this study: input feature > linear transformation (obtaining Query, Key, Value) > dot product to calculate attention weights > normalization (Softmax) > weighted sum > output feature.

[0043] The dimension of the feature map output by the first spatial convolutional layer is H*W*32, that is, 32 spatial feature maps are extracted. The first spatial pooling layer performs max pooling (2*2) on the convolved feature map to reduce the spatial dimension, and the output is H / 2*W / 2*32. The second spatial convolutional layer uses a 3*3 convolutional kernel to extract deeper spatial features, and the output is H / 4*W / 4*64. The second spatial pooling layer performs max pooling again, and the output is H / 8*W / 8*64.

[0044] Exemplarily, the spectral feature extraction branch includes a first spectral convolutional layer, a first spectral pooling layer, and a second spectral convolutional layer. The input of the first spectral convolutional layer is the pixel spectral sequence in the hyperspectral image. A 1*3 convolutional kernel is used to extract spectral features along the spectral dimension, and the dimension of the output feature map is H*W*64, that is, 64 spectral feature maps are extracted. The first spectral pooling layer performs max pooling on the convolved spectral feature map, and the output is H*W*64. The second spectral convolutional layer uses a 1*3 convolutional kernel to continue extracting spectral features, and the output is H*W*128.

[0045] Among them, the number of output channels of the convolutional layers in each branch gradually increases as the network deepens. Therefore, increasing the number of channels can enhance the expression ability of the model. The pooling layer reduces the spatial dimension of the feature map through dimensionality reduction and improves the calculation efficiency.

[0046] In some embodiments of this specification, the feature fusion layer 230 performs feature fusion based on spatial features, spectral features, and dynamic weighting coefficients to obtain a fused feature map, and the dynamic weighting coefficients are adjusted based on the detection accuracy of the model detection module 200.

[0047] Illustratively, the outputs of the spatial feature extraction branch and the spectral feature extraction branch are fused by weighted averaging. The fusion weights of the spatial features and the spectral features are dynamically adjusted according to the detection accuracy of the model. For example, the weights of the spatial features and the spectral features are α and (1 - α) respectively, where α is a hyperparameter reflecting the importance of the spatial features and the spectral features to the final decision. For example, the weight of the spatial features is 0.35 and the weight of the spectral features is 0.65.

[0048] Explanatorily, adopting a dynamic weighting fusion strategy to dynamically adjust the spatial and spectral weights according to task requirements is better than a fixed weight design. By performing adaptive weight analysis, two networks are built together to achieve the purpose of fused diagnosis of atlas data.

[0049] Exemplarily, the method for determining the dynamic weighting coefficients is to use an initial value of α = 0.1, a step size of 0.1, a maximum α = 1, and loop 10 times. Through loop training, the final dynamic weighting coefficients are determined based on the detection accuracies corresponding to the respective values of α.

[0050] In some embodiments of this specification, the adjusted multi-dimensional vector includes a plurality of one-dimensional vectors corresponding one-to-one to feature variables; The second fully connected layer 250 includes a mode selection unit, a first calculation unit, a second calculation unit, and a function unit; The mode selection unit selects the calculation mode of the detection score based on the proportion of the one-dimensional vectors with qualified detection accuracy among all one-dimensional vectors to the total number of qualified one-dimensional vectors. Among them, the detection accuracy of each one-dimensional vector is updated in real time. When the proportion of qualified vectors is greater than a second preset threshold, the calculation mode is the preferred mode; when the proportion of qualified vectors is less than or equal to the second preset threshold, the calculation mode is the global mode; The first calculation unit, in the preferred mode, obtains a first detection score based on the one-dimensional vectors with qualified detection accuracy among all one-dimensional vectors and the weight values in their respective corresponding weight matrices; The second calculation unit, in the global mode, obtains a second detection score based on all one-dimensional vectors and the weight matrix; The function unit obtains the disease probability value of the first detection score or the second detection score according to the Sigmoid activation function.

[0051] Explanatory, each one-dimensional vector can be regarded as a sub-model separately. The method of preferentially selecting multiple sub-models with higher detection accuracy and excluding other sub-models with lower detection accuracy to replace all one-dimensional vectors for the calculation of the detection branch helps to improve the detection accuracy. However, if the number of multiple sub-models with higher detection accuracy selected is small, due to the limitations of each sub-model, it can only show higher accuracy in a specific feature space and cannot globally handle the complexity of the entire task. Therefore, false alarms may occur. The global model can combine the advantages of different sub-models and make more comprehensive judgments in multiple feature spaces. Its goal is to obtain a globally optimal decision by synthesizing the decision results of sub-models. By learning how to synthesize the predictions of different sub-models, the global model can make more robust judgments when facing different data inputs.

[0052] Exemplary, if the detection accuracy of more than half of the sub-models exceeds the set threshold, that is, the second preset threshold is one-half, then a direct determination is made according to the results of the sub-models with more than half. If the number of sub-models with qualified detection accuracy does not exceed the set threshold, a re-determination is made through the weighted average of all one-dimensional vectors and the weight matrix to ensure that the final result has high reliability.

[0053] In some embodiments of the present specification, the user interaction module 300 also outputs the coordinates of the diseased area of the target fish body 400 based on the spatial features and spectral features.

[0054] Explanatory, providing more detailed information about the diseased area is convenient for comparison with the target fish body 400 to verify the detection results.

[0055] In some embodiments of the present specification, the model detection module 200 further includes a model training unit; The model training unit obtains a training data set and trains a model to be trained including a one-dimensional depth convolutional neural network and an image convolutional neural network based on the training data set to obtain a trained deep learning diagnosis model; Among them, the training data set includes multiple training samples corresponding to fish bodies 400 with different disease levels. Each training sample includes hyperspectral image data on the surface of the fish body 400 and the corresponding infection situation of the fish body 400, and the infection situation is determined by PCR detection.

[0056] Explanatory. First, a large amount of hyperspectral image data of channel catfish with different degrees of disease (including symptomatic and asymptomatic) and uninfected channel catfish is collected. The diseased and uninfected catfish are identified through PCR. The spectral information in the training samples is extracted by the one-dimensional depth convolutional neural network in the model to be trained, and the image information in the training samples is extracted by the image convolutional neural network, so as to obtain the key features of disease diagnosis and build a stable deep learning diagnosis model. During the model training process, the collected training samples are divided into a training set, a validation set, and a test set according to a ratio of 2:1:1.

[0057] It can be understood that the detection system 1 can receive the newly uploaded hyperspectral image data from the user in real time and continuously expand the training data set. The newly added hyperspectral image data will further enrich the database, provide more samples for subsequent diagnosis, and improve the generalization ability and accuracy of the model.

[0058] In some embodiments of the present specification, obtaining the training data set in the model training unit includes: Obtain the average value of the data collected three times by the data acquisition module 100 under constant temperature and humidity conditions as the hyperspectral image data in the training samples.

[0059] Explanatory. Ensure the accuracy of the data input into the deep learning model.

[0060] In some embodiments of the present specification, obtaining the training data set in the model training unit further includes: Correct the average value of the data collected three times by the data acquisition module 100 under constant temperature and humidity conditions based on the dark reference image and the white reference image to obtain the hyperspectral image data; Among them, the dark reference image is the image data collected by the CCD camera 115 when the reflectivity is 0, and the white reference image is the image data collected by the CCD camera 115 when the reflectivity is 100.

[0061] Explanatory. Use the dark reference image (Idark, collected by covering the camera lens while turning off the light source) and the white reference image (Iwhite, obtained from a white Teflon tile with a reflectivity of approximately 100%) to correct the original intensity signal to an actual reflectivity image. Calculate the corrected image (Icor) from the original image (Iraw) through the following equation: Explanatory. Preprocess the hyperspectral image data to ensure the consistency of the data input into the deep learning model.

[0062] In some embodiments, preprocessing the hyperspectral image data further includes operations such as denoising, image cropping, normalization, and enhancement.

[0063] In some embodiments of this specification, the model training unit uses the cross-entropy loss function to judge the training situation of the model to be trained.

[0064] Illustratively, the change in the loss value calculated by the cross-entropy loss function is used to evaluate whether the model has learned effective features from the data. When the loss function reaches a minimum value in the parameter space and the gap between the training and validation losses is small, it is judged as ideal convergence. If the training loss is extremely low and the validation loss is high, it indicates that the model has lost its generalization ability and is judged as overfitting convergence; if both the training and validation losses are high, the model cannot learn effective features and is judged as underfitting convergence.

[0065] Explanatorily, in a classification task, the difference between the model's predicted value and the actual label is calculated. The cross-entropy can more directly measure the difference in probability distributions, while the mean squared error (MSE) is more suitable for regression problems. In addition, the gradient of the cross-entropy is more appropriate during the training process, avoiding the problem of vanishing gradients that MSE may bring, especially when using sigmoid or softmax activation functions.

[0066] Effect verification: Using the present detection system 1, the detection accuracy is verified using the trained deep learning diagnosis model. A total of 20 channel catfish are selected, 10 of which are diseased and 10 are healthy. Ten rounds of verification are carried out, and the samples are randomly placed each time for accuracy verification. The test results are shown in Table 2. Round Detection accuracy The 1st round 92.50% The 2nd round 91.00% The 3rd round 93.00% The 4th round 90.50% The 5th round 91.75% The 6th round 94.00% The 7th round 90.00% The 8th round 92.25% The 9th round 93.50% The 10th round 90.75% Average 91.25%

[0067] Table 2: Detection accuracy of the detection system.

[0068] The above-described embodiments are merely described in terms of the preferred embodiments of this specification and do not limit the scope of this specification. Without departing from the design spirit of this specification, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of this specification shall fall within the protection scope determined by the claims of this specification.

Claims

1. A novel acute septicemia detection system for fish based on hyperspectral imaging, characterized in that, It includes a data acquisition module, a model detection module, and a user interaction module: The data acquisition module acquires hyperspectral image data of the surface of the target fish body; The model detection module diagnoses based on the trained deep learning diagnosis model with the hyperspectral image data as the input, including an input layer, a convolutional layer, a feature fusion layer, a first fully connected layer, and a second fully connected layer; The input layer obtains the hyperspectral image data of the surface of the target fish body; The convolutional layer obtains the pixel spatial structure and pixel spectral sequence on the surface of the fish body based on the hyperspectral image data, and identifies the spatial features representing relevant clinical symptoms based on the pixel spatial structure, and identifies the spectral features representing the inositol metabolism situation based on the pixel spectral sequence; The feature fusion layer performs feature fusion based on the spatial features and spectral features to obtain a fused feature map; The first fully connected layer compresses the dimension of the fused feature map to obtain a multi-dimensional vector corresponding to the number of feature variables, and obtains an adjusted multi-dimensional vector corresponding to the multi-dimensional vector according to the ReLU activation function; The second fully connected layer obtains a detection score based on the adjusted multi-dimensional vector and the weight matrix, and obtains a disease probability value corresponding to the detection score according to the Sigmoid activation function; The user interaction module outputs the diagnosis result of the target fish body based on the disease probability value and the first preset threshold.

2. The novel acute septicemia detection system for fish based on hyperspectral imaging according to claim 1, characterized in that, The data acquisition module includes an acquisition unit and a control unit; The acquisition unit includes a black box, a halogen light source arranged inside the black box, a conveyor belt, a CCD camera located above the conveyor belt, and a spectrometer connected to the CCD camera; The control unit is connected to the spectrometer.

3. The novel acute septicemia detection system for fish based on hyperspectral imaging according to claim 1, wherein, The convolutional layer includes a data separation unit, a spatial feature extraction unit, and a spectral feature extraction unit; The data separation unit obtains the pixel spatial structure and pixel spectral sequence on the surface of the fish body based on the hyperspectral image data; The spatial feature extraction unit uses the expert annotation method to obtain clinical symptom recognition labels including punctate hemorrhage, tail fading, and rot, and identifies the spatial features in the pixel spatial structure based on the clinical symptom recognition labels; The spectral feature extraction unit uses the competitive adaptive reweighted sampling method, the successive projection algorithm, and the random frog leaping method to respectively identify the spectral feature wavelengths related to the inositol metabolism situation in the pixel spectral sequence from the preset wavelength range to obtain spectral features.

4. A novel acute septicemia detection system for fish based on hyperspectral imaging according to claim 1, characterized in that, The feature fusion layer performs feature fusion based on the spatial features, spectral features, and dynamic weighting coefficients to obtain a fused feature map, and the dynamic weighting coefficients are adjusted based on the detection accuracy of the model detection module.

5. The novel acute septicemia detection system for fish based on hyperspectral imaging according to claim 1, characterized in that, The adjusted multi-dimensional vector includes multiple one-dimensional vectors corresponding one by one to the feature variables; The second fully connected layer includes a mode selection unit, a first calculation unit, a second calculation unit, and a function unit; The mode selection unit selects the calculation mode of the detection score based on the compliance ratio of the one-dimensional vectors with qualified detection accuracy among all the one-dimensional vectors to all the one-dimensional vectors. Among them, the detection accuracy of each one-dimensional vector is updated in real time. When the compliance ratio is greater than the second preset threshold, the calculation mode is the preferred mode. When the compliance ratio is less than or equal to the second preset threshold, the calculation mode is the global mode; The first calculation unit, in the preferred mode, obtains a first detection score based on the one-dimensional vectors with qualified detection accuracy among all one-dimensional vectors and the weight values in their respective corresponding weight matrices; The second calculation unit, in the global mode, obtains a second detection score based on all one-dimensional vectors and the weight matrix; The function unit obtains the disease probability value of the first detection score or the second detection score according to the Sigmoid activation function.

6. The novel acute septicemia detection system for fish based on hyperspectral imaging according to claim 1, wherein, The user interaction module also outputs the coordinates of the lesion area of the target fish body based on the spatial features and spectral features.

7. A novel acute septicemia detection system for fish based on hyperspectral imaging according to any one of claims 1-6, characterized in that, The model detection module further includes a model training unit; The model training unit obtains a training data set and trains a model to be trained including a one-dimensional depth convolutional neural network and an image convolutional neural network based on the training data set to obtain a trained deep learning diagnosis model; Wherein, the training data set includes a plurality of training samples corresponding to fish bodies with different degrees of infection, and each training sample includes hyperspectral image data on the surface of the fish body and the infection condition corresponding to the fish body, and the infection condition is determined by PCR detection.

8. An innovative acute septicemia detection system for fish based on hyperspectral imaging according to claim 7, characterized in that, Obtaining the training data set in the model training unit includes: Obtaining the average value of the data collected three times by the data acquisition module under constant temperature and humidity conditions as the hyperspectral image data in the training sample.

9. The novel acute septicemia detection system for fish based on hyperspectral imaging according to claim 8, wherein, Obtaining the training data set in the model training unit further includes: Correcting the average value of the data collected three times by the data acquisition module under constant temperature and humidity conditions based on the dark reference image and the white reference image to obtain hyperspectral image data; Wherein, the dark reference image is the image data collected by the CCD camera when the reflectivity is 0, and the white reference image is the image data collected by the CCD camera when the reflectivity is 100.

10. A novel acute septicemia detection system for fish based on hyperspectral imaging according to claim 7, wherein: The model training unit uses the cross-entropy loss function to judge the training situation of the model to be trained.

Images