An underwater optical harmful organism recognition method and device and equipment

Abstract

The application discloses an underwater optical disaster-causing biological identification method, device and equipment. The method comprises the following steps: obtaining different underwater picture data, and marking disaster-causing organisms; constructing and training a teacher neural network based on the underwater picture data; generating a student neural network through the teacher neural network, and training the student neural network through the output result of the teacher neural network; and predicting and classifying the disaster-causing organisms corresponding to the underwater picture data according to the student neural network. The application uses the lightweight student YOLO network to realize the portability of equipment operation. The technical problem that small devices are difficult to deploy strong hardware to meet the performance requirements of complex networks is solved. The multi-dimensional feature extraction method makes the network more easily identify objects of different scales. The detection head of the network structure is modified, so that each pixel point of the feature map in the prediction process corresponds to at least two prediction results, and the recognition rate when underwater organisms overlap is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of age prediction, in particular to an underwater optical disaster-causing organism identification method and device and equipment. BACKGROUND

[0002] Nuclear power relies on the heat energy generated by nuclear reactions to convert into electric energy to generate electricity, which requires a large amount of cooling water to ensure the safe operation of nuclear power, so nuclear power plants are generally built in coastal areas. There are many nuclear power plants along the coast of China. The nuclear power intake is prone to invasion by organisms (brine shrimp, jellyfish, sea cucumbers, fish, etc.) that erupt in the sea, causing blockage, and further causing nuclear power unit shutdown, resulting in economic losses and affecting the safe operation of nuclear power. Therefore, a system needs to be designed to monitor the nuclear power intake in real time and estimate the biomass to ensure the safe operation of the nuclear power plant.

[0003] The main method of identifying targets is to use an optical camera to take pictures, process the pictures, and use a neural network to identify the existing targets. The related identification technology can be used for underwater optical identification for monitoring disaster-causing organisms. However, the land (such as pedestrian, vehicle, and other object recognition) and the underwater environment are different, which is mainly reflected in that the underwater environment is complex and requires small devices to dive and shoot, and it is difficult to deploy strong hardware for small devices to meet the performance requirements of complex networks. The individual differences of underwater disaster-causing organisms are large, such as brine shrimp, which is generally 3-4 cm in size, small and flat, and sea cucumbers, which are generally 10-20 cm in size, similar in shape and color to sweet potatoes. The existing network directly applies to the case where small and large volume organisms overlap, and the recognition rate is low. Therefore, the applicant proposes an underwater optical disaster-causing organism identification method and device and equipment. SUMMARY

[0004] Therefore, the purpose of the present application is to provide an underwater optical disaster-causing organism identification method and device and equipment, which can realize the prediction and classification of disaster-causing organisms through underwater picture data.

[0005] According to one aspect of the present application, an underwater optical disaster-causing organism identification method is provided, comprising:

[0006] Obtaining different underwater picture data and labeling disaster-causing organisms;

[0007] Constructing and training a teacher neural network based on the underwater picture data, the teacher neural network structure comprising a feature extraction network and a detection head, the feature extraction network performing feature extraction on the labeled underwater picture data and outputting at least three-dimensional feature maps; inputting the at least three-dimensional feature maps into the detection head for prediction, each pixel point of the feature map corresponding to at least two prediction results in the prediction process;

[0008] The student neural network is generated by the teacher neural network, and the student neural network is trained by the output result of the teacher neural network; wherein the architecture of the student neural network comprises the structure and parameters of the teacher neural network, and the number of convolution kernels of the student neural network is 25% of the teacher neural network.

[0009] The underwater picture data is predicted and classified according to the student neural network.

[0010] According to another aspect of the present application, an underwater optical disaster-causing organism identification device is provided, comprising an acquisition module, a teacher network module, and a student network module.

[0011] The acquisition module is configured to acquire different underwater picture data and label disaster-causing organisms.

[0012] The teacher network module is configured to construct and train a teacher neural network based on the underwater picture data, wherein the structure of the teacher neural network comprises a feature extraction network and a detection head, the feature extraction network performs feature extraction on the labeled underwater picture data and outputs feature maps of at least three dimensions, and the feature maps of the at least three dimensions are input into the detection head for prediction, wherein each pixel point of the feature maps corresponds to at least two prediction results in the prediction process.

[0013] The student network module is configured to generate a student neural network by the teacher neural network, and train the student neural network by the output result of the teacher neural network; wherein the architecture of the student neural network comprises the structure and parameters of the teacher neural network, and the number of convolution kernels of the student neural network is 25% of the teacher neural network.

[0014] The deployment module is configured to deploy the student network, and predict and classify disaster-causing organisms corresponding to the picture data according to the student neural network.

[0015] According to still another aspect of the present application, an underwater optical disaster-causing organism identification device is provided, comprising at least one processor and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the underwater optical disaster-causing organism identification method according to any one of the above.

[0016] According to still another aspect of the present application, a computer readable storage medium is provided, which stores a computer program, wherein the computer program is executed by a processor to implement the underwater optical disaster-causing organism identification method according to any one of the above.

[0017] It can be found that the above scheme trains a lightweight student YOLO network through a teacher YOLO network, and realizes the portability of device operation by using the lightweight student YOLO network, so that the designed early warning system has the characteristics of low cost and mobility. The technical problem that small devices are difficult to deploy strong hardware to meet the performance requirements of complex networks is solved. For the problem that the individual differences of underwater disaster-causing organisms are large, the present application adopts a multi-dimensional feature extraction method for underwater picture data. Because small-sized objects can only extract obvious features under low-dimensional feature maps, large-sized objects can only extract local features. After multiple extractions, high-dimensional global features are obtained. Small-sized objects are prone to lose features, and large-sized objects can extract their global features. Therefore, the fusion of three different dimensional feature maps can make the network more easily identify objects of different scales. For the technical problem that the recognition rate of small-sized organisms and large-sized organisms after overlapping is low, the present application modifies the detection head of the network structure, so that each pixel point of the feature map corresponds to at least two prediction results in the prediction process, thereby improving the recognition rate when underwater organisms overlap. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0019] Figure 1 is a flowchart of an embodiment of the underwater optical disaster-causing organism recognition method of the present application;

[0020] Figure 2 is a YOLOV5 teacher network structure diagram of an embodiment of the underwater optical disaster-causing organism recognition method of the present application;

[0021] Figure 3 is a YOLOV5 student network structure diagram of an embodiment of the underwater optical disaster-causing organism recognition method of the present application;

[0022] Figure 4 is an output structure diagram of the YOLOV5 teacher network and the YOLOV5 student network of an embodiment of the underwater optical disaster-causing organism recognition method of the present application. DETAILED DESCRIPTION

[0023] The application will be described in further detail below with reference to the drawings and embodiments. It is particularly pointed out that the following embodiments are only for illustrating the application, but not for limiting the scope of the application. Similarly, the following embodiments are only part of the embodiments of the application, but not all the embodiments. All other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the application.

[0024] The application provides an underwater optical disaster-causing organism identification method, which can realize prediction and classification of disaster-causing organisms through corresponding picture data of underwater picture data.

[0025] Please refer to Figure 1 , Figure 1 is a flowchart of an embodiment of the underwater optical disaster-causing organism identification method of the application. It should be noted that the method of the application is not limited to the flow sequence shown in Figure 1 . As shown in Figure 1 , the method comprises the following steps:

[0026] S101, acquiring different underwater picture data and labeling disaster-causing organisms;

[0027] In this embodiment, the data set for training is pictures of shrimps and jellyfish searched on the Internet and pictures of brine shrimps and sea cucumbers cropped from horizontal videos taken in the sea area near the nuclear power plant. The pictures of sea cucumbers are captured in the field. The collected pictures are randomly rotated to increase the number of pictures. Before training the YOLOV5 network, the pictures need to be labeled in advance. The picture labeling format used in the application is VOC format. The position of each target in each picture is manually framed and labeled in advance. The VOC format records the coordinate values of the upper left and lower right of each target frame. After labeling each picture, the data set is completed. 80% of the finally obtained data set is divided into a training set for training the YOLOV5 neural network, 10% is used as a test set for testing the error of the network, and the last 10% is used as a validation set to verify the effectiveness of the training.

[0028] S102, constructing and training a teacher neural network based on the underwater picture data;

[0029] In the embodiment, the teacher neural network structure includes a feature extraction network and a detection head, the feature extraction network performs feature extraction on the labeled underwater picture data and outputs high, medium and low three-dimensional feature maps; the three-dimensional feature maps are input into the detection head for prediction, and each pixel point of the feature map corresponds to at least two prediction results in the prediction process; the purpose of such setting is to solve the problem of large individual difference of underwater disaster-causing organisms, and the present application adopts a multi-dimensional feature extraction mode for underwater picture data, because small-sized objects can only be extracted in low-dimensional feature maps, large-sized objects can only be extracted in local features, and after multiple extractions, high-dimensional global features are obtained, small-sized objects are easy to lose features, and large-sized objects can extract global features, so fusion of three different dimensional feature maps can make the network more easily identify objects of different scales. The problem of low recognition rate after the overlap of small-sized organisms and large-sized organisms is solved, the detection head of the network structure is modified, so that each pixel point of the feature map corresponds to at least two prediction results in the prediction process, and the recognition rate when the underwater organisms overlap is improved.

[0030] S103, generating a student neural network through the teacher neural network, training the student neural network through the output result of the teacher neural network;

[0031] In the embodiment, the architecture of the student neural network includes the structure and parameters of the teacher neural network, and the number of convolution kernels of the student neural network is 25% of that of the teacher neural network.

[0032] In the embodiment, the network structure of the teacher network is as shown in Figure 2 ; the student network is a small and lightweight network actually deployed, and the network structure thereof is as shown in Figure 3 ; regardless of the teacher network or the student network, the final output size is the same, the width of the backbone part of the student network is 25% of that of the teacher network, so as to realize the lightweight of the network. YOLOV5 predicts each pixel point of the final three outputs, and since one pixel point may correspond to multiple targets of the original picture in a dense scene, if the target class to be identified is n, the output structure is as shown in Figure 4 ;

[0033] The teacher network and the student network both have a feature extraction part and a detection head part. The feature extraction part mainly extracts pixel information on the image by convolution kernels. After multiple extractions, high-dimensional information of the image, i.e., feature maps, can be obtained, and the size of the image is reduced to facilitate the processing of the detection head. In order to ensure that different size targets can be detected as much as possible, the feature extraction network outputs three different dimensional feature maps. The detection head of YOLOV5 predicts whether each pixel point of the feature map exists, the category of the target, and the recognition box coordinates of the target. The final output of the network is a set of vectors:

[0034] (conf,class,x,y,w,h)

[0035] where conf is the confidence, indicating the reliability of the recognition box, class is the category of the target, x is the horizontal coordinate of the center point of the recognition box, y is the vertical coordinate of the center point of the recognition box, w is the width of the center point of the recognition box, and h is the length of the center point of the recognition box. After training, only the recognition box center length and width are needed to draw a rectangular prediction box in the original image to frame the recognized target, and the category of the prediction box is output according to the value of class.

[0036] Then, the GIoU of each recognition box in the output vector set and the real labeled recognition box is calculated, and the recognition box with GIoU higher than 0.5 is finally output as the final output result. The calculation method of GIoU is shown in the following figure, where A and B represent two rectangles, and C represents the smallest rectangle containing A and B.

[0037] In the training stage, the error of the output recognition box needs to be calculated, so the error between the output recognition box and the real box is calculated, and the calculation formula of the loss function is as follows:

[0038] BCELoss(x,y)=-w(y log x+(1-y)log(1-x))

[0039]

[0040]

[0041]

[0042] loss=loss conf +loss loc +loss cls

[0043] where conf n represents the confidence of the nth prediction, gt n represents whether the nth object exists, and loc represents the center point and length and width of the prediction box, gt_locn represents the center point and the length-width of the nth real frame, cls represents the predicted class, gt cls n represents the real class of the nth, loss conf represents the confidence error, loss loc represents the bounding box error, loss cls represents the classification error. The total loss is the sum of the above confidence error, bounding box error and classification error.

[0044] By back-propagating the error results, all parameters of the network can be updated. With the increase of the number of iterations, the error will gradually decrease, which means that the network has learned the features of the target. The weight is saved once per iteration. In actual deployment and application, only the saved weight file needs to be placed on the actual deployment host to be used. For example, after training the student network, the saved weight of the student network is placed on the small host to run to realize recognition.

[0045] The teacher network, as a large network, has stronger learning ability than the student network, so the teacher network needs to be trained first to obtain the best recognition result. Then, the student network is trained to recognize the ability in the way of knowledge distillation. Compared with directly learning the real label, it is easier to learn the softened prediction label of the teacher network. Therefore, a temperature variable T is introduced to soften the output of the teacher network. The student network calculates the error between the softened output of the teacher network and the real label output, so that a smaller network structure can learn stronger recognition ability.

[0046]

[0047] The above formula is the distillation target, where x_i represents the probability value of one of the predicted classes, T represents the temperature, and N represents the total number of predicted classes. By softening the original output result through T, the learning efficiency of the student network can be improved. After multiple rounds of training, the weight of the student network is recorded. When actually outputting, only T=1 is needed to restore the original prediction result that has not been distilled. The loss function of the student network is as follows:

[0048]

[0049] loss gt =loss conf +loss loc +loss cls

[0050] loss teacher =L soft (conf t , conf s ) + L soft (loct , loc s )+L soft (cls t , cls s )

[0051] loss=α*loss teacher +β*loss gt

[0052] where conf t is the confidence output of the teacher network, conf s is the confidence output of the student network, loc t is the bounding box output of the teacher network, loc s is the bounding box output of the student network, cls t is the predicted class output of the teacher network, cls s is the predicted class output of the student network, loss gt refers to the error of the student network relative to the true value, loss teacher refers to the error of the student network relative to the output value of the teacher network, and alpha and beta are experimental parameters.

[0053] S104, according to the student neural network, the underwater picture data is corresponding to the prediction and classification of the disaster-causing organisms of the picture data;

[0054] The student network weight file trained is saved to a small host, the student network is installed and deployed on the small host, and the small host is light in volume and suitable for installation on a fishing ship. The camera is connected with the small host through a network cable, the pictures shot are transmitted to the small host, after the small host receives the image information, the image information is sent to the student network to extract features, and the target in the current picture is recognized. The recognized target and quantity are marked on the UI interface displayed on the back end, and the staff can judge whether the disaster-causing organisms need to be fished according to the recognized target quantity.

[0055] It can be found that, in the embodiment, the lightweight student YOLO network is trained by the teacher YOLO network, and the portable device operation is realized by using the lightweight student YOLO network, so that the designed early warning system has the characteristics of low cost and mobility. The technical problem that small devices are difficult to deploy strong hardware to meet the performance requirements of complex networks is solved. For the problem that the individual differences of underwater disaster-causing organisms are large, the present application adopts a multi-dimensional feature extraction method for underwater picture data. Because small-sized objects can only extract obvious features in low-dimensional feature maps, large-sized objects can only extract local features. After multiple extractions, high-dimensional global features are obtained. Small-sized objects are prone to lose features, and large-sized objects can extract their global features. Therefore, the fusion of three different dimensional feature maps can make the network more easily identify objects of different scales. For the technical problem that the recognition rate of small-sized organisms and large-sized organisms after overlapping is low, the present application modifies the detection head of the network structure, so that each pixel point of the feature map corresponds to at least two prediction results in the prediction process, thereby improving the recognition rate when underwater organisms overlap.

[0056] The present application also provides an underwater optical disaster-causing organism identification device. In the embodiment, the underwater optical disaster-causing organism identification device comprises an acquisition module, a teacher network module and a student network module.

[0057] The acquisition module is configured to acquire different underwater picture data and label disaster-causing organisms.

[0058] The teacher network module is configured to construct and train a teacher neural network based on the underwater picture data. The teacher neural network structure comprises a feature extraction network and a detection head. The feature extraction network performs feature extraction on the labeled underwater picture data and outputs at least three dimensional feature maps. The at least three dimensional feature maps are input into the detection head for prediction. Each pixel point of the feature map corresponds to at least two prediction results in the prediction process.

[0059] The student network module is configured to generate a student neural network through the teacher neural network and train the student neural network through the output result of the teacher neural network. The architecture of the student neural network comprises the structure and parameters of the teacher neural network. The number of convolution kernels of the student neural network is 25% of that of the teacher neural network.

[0060] The deployment module is configured to deploy the student network and perform prediction and classification of disaster-causing organisms corresponding to the picture data according to the student neural network.

[0061] Optionally, the teacher network module can be specifically configured to:

[0062] The underwater picture data is data augmented and labeled, and the YOLOV5 network is used to extract pixel information from the labeled underwater picture data multiple times to obtain a feature map of each underwater picture data, and output three-dimensional feature maps as the training input of the YOLOV5 network; and the YOLOV5 network finally outputs a vector:

[0063] (conf, class, x, y, w, h)

[0064] Wherein conf is the confidence, indicating the reliability of the recognition box, class is the category of the target, x represents the horizontal coordinate of the center point of the recognition box, y represents the vertical coordinate of the center point of the recognition box, w represents the width of the center point of the recognition box, and h represents the length of the center point of the recognition box.

[0065] Then, the GIoU of each recognition box in the output vector group and the real labeled recognition box is calculated, and the recognition box with GIoU higher than 0.5 is finally output as the final output result.

[0066] Optionally, the student network module can be specifically used for:

[0067] According to the trained teacher network, the prediction label output by the teacher network is softened, and the student network is used to learn the softened prediction label, and the error between the output of the student network and the output of the real label after the teacher network is softened is calculated. The loss function of the student network is as follows:

[0068]

[0069] loss gt =loss conf +loss loc +loss cls

[0070] loss teacher =L soft (conf t ,conf s )+L soft (loc t ,loc s )+L soft (cls t ,cls s )

[0071] loss=α*loss teacher +β*loss gt

[0072] Wherein conf t is the confidence output of the teacher network, conf sis the confidence output of the student network, loc t is the predicted bounding box output of the teacher network, loc s is the predicted bounding box output of the student network, cls t is the predicted class output of the teacher network, cls s is the predicted class output of the student network, loss gt refers to the error of the student network relative to the true value, loss teacher refers to the error of the student network relative to the output value of the teacher network, and alpha and beta are experimental parameters.

[0073] The underwater optical disaster-causing organism identification device also includes an updating module.

[0074] The updating module is configured to calculate a final error based on the error of the teacher network and the error of the student network, perform parameter updating on the student network based on the error result by back propagation, and train and update the student network by iteration of a prediction number of times using the student network after the parameter updating.

[0075] The various unit modules of the underwater optical disaster-causing organism identification device can respectively perform the corresponding steps in the method embodiments described above, and thus the unit modules will not be described in detail here. For details, please refer to the description of the corresponding steps above.

[0076] The present application also provides an underwater optical disaster-causing organism identification device, comprising: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the underwater optical disaster-causing organism identification method described above.

[0077] The memory and the processor are connected in a bus manner, the bus can include any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors and memories together. The bus can also connect various other circuits such as peripheral devices, voltage stabilizers, and power management circuits together, which are well known in the art, and thus they will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be one element or multiple elements such as multiple receivers and transmitters, which provide a unit for communicating with various other devices on a transmission medium. Data processed by the processor is transmitted on a wireless medium through an antenna, and further, the antenna also receives data and transmits the data to the processor.

[0078] The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interface, voltage regulation, power management, and other control functions. The memory can be used to store data used by the processor in performing operations.

[0079] The application further provides a computer readable storage medium storing a computer program. The computer program is executed by the processor to implement the method embodiments.

[0080] It can be found that the above scheme trains a lightweight student YOLO network through a teacher YOLO network, and realizes the portability of device operation by using the lightweight student YOLO network, so that the designed early warning system has the characteristics of low cost and mobility. The technical problem that small devices are difficult to deploy strong hardware to meet the performance requirements of complex networks is solved. For the problem that the individual differences of underwater disaster-causing organisms are large, the application adopts a multi-dimensional feature extraction method for underwater picture data. Because small-sized objects can only extract obvious features in low-dimensional feature maps, large-sized objects can only extract local features. After multiple extractions, high-dimensional global features are obtained. Small-sized objects are prone to lose features, and large-sized objects can extract their global features. Therefore, the fusion of three different dimensional feature maps can make the network more easily identify objects of different scales. For the technical problem that the recognition rate is low when small-sized organisms and large-sized organisms overlap, the application modifies the detection head of the network structure, so that each pixel point of the feature map corresponds to at least two prediction results in the prediction process, thereby improving the recognition rate when underwater organisms overlap.

[0081] In several embodiments provided by the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic, and the division of the modules or units is only a logical function division, and there can be another division manner in actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0082] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the present embodiment scheme.

[0083] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0084] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0085] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An underwater optical biometric identification method for disasters, characterized in that, include: Acquire different underwater image data and label the organisms that caused the disaster; A teacher neural network based on the underwater image data is constructed and trained. The teacher neural network structure includes a feature extraction network and a detection head. The feature extraction network extracts features from the labeled underwater image data and outputs feature maps of at least three dimensions. The feature maps of at least three dimensions are input into the detection head for prediction. During the prediction process, each pixel of the feature map corresponds to at least two prediction results. A student neural network is generated by a teacher neural network, and the student neural network is trained using the output of the teacher neural network; wherein the architecture of the student neural network includes the structure and parameters of the teacher neural network, and the number of convolutional kernels in the student neural network is 25% of that in the teacher neural network; The student neural network is used to predict and classify the disaster-causing organisms corresponding to the underwater image data. The construction and training of the teacher neural network based on the underwater image data includes: The underwater image data is augmented and labeled, and the YOLOv5 network is used to extract pixel information multiple times from the labeled underwater image data to obtain feature maps for each underwater image data. Three-dimensional feature maps are output as training inputs to the YOLOv5 network; and the final output vector of the YOLOv5 network is as follows: Where conf is the confidence level, representing the confidence level of the bounding box; class is the category of the target; x is the x-coordinate of the center point of the bounding box; y is the y-coordinate of the center point of the bounding box; w is the width of the center point of the bounding box; and h is the length of the center point of the bounding box. Then, the GIoU is calculated between each bounding box in the output vector group and the ground truth bounding box. Finally, the bounding boxes with a GIoU higher than 0.5 are output as the final output result. The process of generating a student neural network through a teacher neural network and training the student neural network using the output of the teacher neural network includes: Based on the trained teacher network, its predicted labels are softened. The softened predicted labels are then used to train the student network. The error between the softened output of the teacher network and the true label output is calculated from the student network's output. The loss function of the student network is as follows: in It is the confidence output of the teacher network. It is the confidence output of the student network. It is the predicted bounding box output of the teacher network. It is the predicted bounding box output of the student network. It is the predicted category output of the teacher network. It is the predicted category output of the student network. This refers to the error of the student network relative to the actual value. This refers to the error between the student network's output value and the teacher network's output value. and These are experimental parameters.

2. The underwater optical disaster-causing biometric identification method as described in claim 1, characterized in that, The process of generating a student neural network through a teacher neural network and training the student neural network using the output of the teacher neural network further includes: The final error is calculated by combining the errors of the teacher network and the student network. Based on this error result, backpropagation is performed to update the parameters of the student network. The student network is then trained and updated using the updated parameters through iterative prediction.

3. An underwater optical biometric identification device for causing disasters, characterized in that, Based on the method according to any one of claims 1-2, the apparatus comprises: Acquisition module, teacher network module, student network module, deployment module; The acquisition module is used to acquire different underwater image data and label the organisms that caused the disaster. The teacher network module is used to construct and train a teacher neural network based on the underwater image data. The teacher neural network structure includes a feature extraction network and a detection head. The feature extraction network extracts features from the labeled underwater image data and outputs feature maps of at least three dimensions. The feature maps of at least three dimensions are input into the detection head for prediction. During the prediction process, each pixel of the feature map corresponds to at least two prediction results. The student network module is used to generate a student neural network from the teacher neural network and to train the student neural network using the output of the teacher neural network; wherein the architecture of the student neural network includes the structure and parameters of the teacher neural network, and the number of convolutional kernels in the student neural network is 25% of that in the teacher neural network; The deployment module is used to deploy a student network and, based on the student neural network, predict and classify the disaster-causing organisms corresponding to the underwater image data.

4. The underwater optical biometric identification device for causing disaster as described in claim 3, characterized in that, The teacher network module is specifically used for: The underwater image data is augmented and labeled, and the YOLOv5 network is used to extract pixel information multiple times from the labeled underwater image data to obtain feature maps for each underwater image data. Three-dimensional feature maps are output as training inputs to the YOLOv5 network; and the final output vector of the YOLOv5 network is as follows: Where conf is the confidence level, representing the confidence level of the bounding box; class is the category of the target; x is the x-coordinate of the center point of the bounding box; y is the y-coordinate of the center point of the bounding box; w is the width of the center point of the bounding box; and h is the length of the center point of the bounding box. Then, the GIoU is calculated between each bounding box in the output vector group and the ground truth bounding box. Finally, the bounding boxes with a GIoU higher than 0.5 are output as the final result.

5. The underwater optical biometric identification device for causing disaster as described in claim 3, characterized in that, The student network module is specifically used for: Based on the trained teacher network, its predicted labels are softened. The softened predicted labels are then used to train the student network. The error between the softened output of the teacher network and the true label output is calculated from the student network's output. The loss function of the student network is as follows: in It is the confidence output of the teacher network. It is the confidence output of the student network. It is the predicted bounding box output of the teacher network. It is the predicted bounding box output of the student network. It is the predicted category output of the teacher network. It is the predicted category output of the student network. This refers to the error of the student network relative to the actual value. This refers to the error between the student network's output value and the teacher network's output value. and These are experimental parameters.

6. The underwater optical biometric identification device for causing disaster as described in claim 3, characterized in that, The underwater optical biometric identification device also includes: Update module; The update module is used to calculate the final error through the error of the teacher network and the error of the student network, perform backpropagation based on the error result to update the parameters of the student network, and use the updated parameters to train and update the student network through the number of prediction iterations.

7. An underwater optical biometric identification device for causing disasters, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the underwater optical disaster biometric identification method as described in any one of claims 1 to 2.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the underwater optical disaster-causing biometric identification method as described in any one of claims 1 to 2.

Images