A low-quality underwater image fish target detection method

Abstract

The application discloses a low-quality underwater image fish target detection method, comprising the following steps: collecting an underwater image dataset, pre-processing the image dataset to obtain a model dataset; constructing a GCANet network model, constructing an HF-GCANet defogging network model based on the GCANet network model; constructing a GCA-YOLOv4 target detection network model based on the HF-GCANet defogging network model and in combination with a YOLOv4 network; training the GCA-YOLOv4 target detection network model based on the model dataset; and detecting and classifying a to-be-recognized image based on the trained GCA-YOLOv4 target detection network model to obtain prediction frame position information and classification information. The application has strong universality and can realize fish target detection of more types through expansion of a dataset for training.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of marine resource exploration, and in particular to a low-quality underwater image fish target detection method. BACKGROUND

[0002] The coastal zone is a valuable land resource for coastal countries and a base for marine development and economic development. Accurate extraction of the coastline and analysis of the changes in the coastline provide decision support for government departments to strengthen the management and governance of the coastal zone, which is of great significance for the sustainable development of coastal resources and economic construction in coastal areas. Fisheries have become an important industry for promoting the prosperity of China's rural economy. According to statistics, the total fishery output value in 2019 was 12934.49 billion yuan, the fishery population reached 1828.20 million people, and the total number of fishing vessels was 73.12 million. However, with the large-scale development of fisheries, various problems have emerged. Freshwater aquaculture consumes more and more water resources and occupies more and more land resources. The overuse of resources leads to an increase in the unit price of resources, thereby increasing the cost of aquaculture. The unit area output of the past extensive and single aquaculture method cannot meet the expected income, which to some extent limits the development process of the aquaculture industry. Under the constraints of resource scarcity, China's aquaculture industry has accumulated certain production contradictions. Simply expanding the production scale cannot meet the long-term sustainable development of fisheries. Therefore, how to make full use of material resources is a consideration for the development of fisheries. Therefore, more attention should be paid to the rationalization and standardization of aquaculture, and modern high-tech should be fully utilized to provide basis and support for efficient aquaculture methods.

[0003] Currently, the methods applied to image target detection mainly include two categories: two-stage target detection and one-stage target detection. The two-stage target detection network divides target detection into two steps: first, input the image into the candidate frame recommendation network (RPN) to obtain several candidate frame outputs, and then input the information in the candidate frame region to the convolutional neural network for regression to obtain the classification output. The one-stage target detection network only needs to input the image into the main neural network once during target detection to obtain the category information and position information of the predicted target, without generating candidate frame information through the candidate frame recommendation network (RPN). The algorithm has a lower time complexity and higher real-time performance, but at the same time, it sacrifices part of the accuracy.

[0004] Since underwater fish target detection is usually real-time video image target detection, it requires high real-time performance, so more one-stage target detection networks are used. At present, one-stage target detection networks are rapidly developing, including new algorithms such as SSD and YOLOv4, which have greatly improved detection accuracy and speed, and have become the mainstream method of image target detection.

[0005] In recent years, scholars at home and abroad have continuously researched and made certain achievements in underwater image fish target detection. Ahsan Jalal and others from Pakistan National University of Sciences and Technology combined optical flow and Gaussian mixture model with YOLO network model in 2020, which improved the detection accuracy of fish camouflaged in the background; Kristian and others from Agder University proposed a fish identification method combining CNN-SE (convolutional neural network integrating compression and excitation network) module with YOLO model in 2021, which improved the detection accuracy of temperate fish; Li Qingzhong and others from Ocean University of China added feature fusion to the YOLO network model in 2019, which improved the underwater fish target detection accuracy.

[0006] The above research results at home and abroad only improve the target detection network model from the aspect of image feature information extraction, lack of original image data enhancement processing, which will lead to lower detection accuracy of low-quality underwater images than high-quality underwater images. SUMMARY

[0007] In view of the problem that the existing image detection method mainly improves the image feature fusion, which has obvious effect on high-quality images, but poor effect and low accuracy on low-quality underwater images, the present application proposes a low-quality underwater image fish target detection method, which is based on the image dehazing network GCANet, uses hybrid dilated convolution and feature pyramid network to improve and proposes a new image dehazing network HF-GCANet, uses the dehazing enhancement effect of HF-GCANet on low-quality underwater images, and integrates HF-GCANet into YOLOv4 to design a fish target detection method applied to low-quality underwater images.

[0008] To achieve the above purpose, the present application provides the following scheme: a low-quality underwater image fish target detection method, comprising the following steps:

[0009] S1, collecting underwater image data set, pre-processing the image data set to obtain model data set;

[0010] S2, constructing GCANet network model, constructing HF-GCANet dehazing network model based on the GCANet network model;

[0011] S3, constructing GCA-YOLOv4 target detection network model based on the HF-GCANet dehazing network model and combining YOLOv4 network;

[0012] S4, training the GCA-YOLOv4 target detection network model based on the model data set;

[0013] S5, detecting and classifying the to-be-identified image based on the trained GCA-YOLOv4 target detection network model to obtain prediction frame position information and classification information.

[0014] Preferably, the method for obtaining the model data set in S1 comprises:

[0015] Collecting small fish videos in the preset data, extracting each frame in the videos, converting into image form, and obtaining the image data set;

[0016] Naming the image data set according to time period and frame number and converting into YOLO format to obtain the model data set.

[0017] Preferably, the GCANet network model in S2 comprises an auto-encoder, a smooth expansion module, and a fusion gate fusion subnetwork.

[0018] The auto-encoder is composed of three convolutional layers, the number of channels of the convolutional layers is 64, and the size of the convolution kernel is 3*3;

[0019] The last convolutional layer of the auto-encoder adopts 1 / 2 stride downsampling;

[0020] The smooth expansion module is composed of six smooth hollow residual units and one residual unit.

[0021] The fusion gate fusion subnetwork is composed of one gate fusion subnetwork, one transpose convolutional layer, and two convolutional layers.

[0022] Preferably, the method for constructing the HF-GCANet defogging network model in S2 comprises:

[0023] Hybrid hollow convolution is adopted instead of the smooth hollow convolution in the GCANet network model;

[0024] A feature pyramid network is adopted instead of the fusion gate fusion subnetwork in the GCANet network model.

[0025] Preferably, the hybrid hollow convolution comprises an inverse convolutional layer 1, an inverse convolutional layer 2, an inverse convolutional layer 3, an inverse convolutional layer 4, an inverse convolutional layer 5, and an inverse convolutional layer 6.

[0026] The input and output channel numbers of the inverse convolutional layer 1, the inverse convolutional layer 2, and the inverse convolutional layer 3 are all 64, the stride is 1, and the dilated rate is 1, 2, and 3, respectively;

[0027] The input channel number of the inverse convolutional layer 4 is 64, the output channel number is 128, the stride is 2, and the dilated rate is 1;

[0028] The input channel number of the deconvolution layer 5 is 128, the output channel number is 256, the stride is 2, and the dilated rate is 2.

[0029] The input channel number of the deconvolution layer 6 is 256, the output channel number is 512, the stride is 2, and the dilated rate is 3.

[0030] Preferably, the S3 YOLOv4 network comprises an input module, a backbone network module, a feature extraction module and an output module.

[0031] The input module comprises a Mosaic unit, a CmBN unit and a SAT unit.

[0032] The backbone network module comprises a CBM unit and a CSP unit.

[0033] The feature extraction module comprises an SPP unit and an FPN unit.

[0034] Preferably, the S3 GCA-YOLOv4 target detection network model comprises an input module, an HF-GCA module, a backbone network module, a feature extraction module and an output module.

[0035] The HF-GCA module comprises an auto-encoder, a mixed dilated convolution unit and a feature pyramid network unit.

[0036] Preferably, the method for obtaining the prediction box position information and the classification information comprises the following steps.

[0037] Step 1: input the to-be-recognized image into the input module for image enhancement to obtain an enhanced image.

[0038] Step 2: input the enhanced image into the HF-GCA module to obtain a defogging image.

[0039] Step 3: input the defogging image into the backbone network module for feature extraction to obtain a feature map.

[0040] Step 4: input the feature map into the feature extraction module for feature fusion to obtain a fusion feature vector.

[0041] Step 5: input the fusion feature vector into the output module to obtain the prediction box position information and the classification information.

[0042] The present application has the following advantages.

[0043] Compared with other mainstream target detection, the application fuses GCANet and YOLOv4 to realize fish target detection of low-quality underwater images, which has important significance for exploration and sustainable development of marine fishery resources; the fish target detection method of the low-quality underwater image has strong universality, and can realize detection of more kinds of fish targets by expanding the data set training. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the present application, the drawings needed in the embodiments are briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0045] Figure 1 The step flowchart of the present application is shown in the figure;

[0046] Figure 2 The GCANet network model structure of the present application is shown in the figure;

[0047] Figure 3 The feature pyramid network structure of the present application is shown in the figure;

[0048] Figure 4 The HF-GCANet defogging network model structure of the present application is shown in the figure;

[0049] Figure 5 The mixed hollow convolution structure of the present application is shown in the figure;

[0050] Figure 6 The FPN unit structure of the present application is shown in the figure;

[0051] Figure 7 The YOLOv4 network structure of the present application is shown in the figure;

[0052] Figure 8 The GCA-YOLOv4 target detection network model structure of the present application is shown in the figure;

[0053] Figure 9 Some image data in the model data set of the present application is shown in the figure;

[0054] Figure 10 The target detection result of the GCA-YOLOv4 target detection network model of the present application is shown in the figure. DETAILED DESCRIPTION

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Example 1

[0058] like Figure 1 The diagram shown is a flowchart of a method for detecting fish targets in low-quality underwater images according to the present invention. The method includes the following steps:

[0059] S1. Collect underwater image datasets, preprocess the image datasets to obtain model datasets;

[0060] This embodiment uses videos of small fish from the publicly available European underwater image dataset, The Brackish Dataset. Each frame is extracted using ffmpeg, resulting in 3290 images. Each image is named according to its time segment and frame number. The annotation information for each image is then matched against the existing annotation files in the dataset based on its name and converted to YOLO annotation format to obtain the model dataset. For example... Figure 9 As shown, this is a portion of the image data in the model dataset; finally, the model dataset is divided into training set, validation set, and test set according to the proportions of 80%, 10%, and 10%.

[0061] S2. Construct the GACNet network model, and based on the GCANet network model, construct the HF-GCANet defogging network model;

[0062] like Figure 2 The diagram shown is a schematic of the GCANet network model structure. The GCANet network model includes: an autoencoder, a smooth expansion module, and a gating fusion subnetwork.

[0063] The autoencoder consists of three convolutional layers that encode the input foggy image into a feature image. The last convolutional layer uses a 1 / 2 stride downsampling to increase the receptive field and reduce the computational cost of feature extraction. All three convolutional layers have 64 channels and a 3×3 kernel size, allowing the autoencoder to obtain more image feature information while reducing computation time.

[0064] The smooth expansion module is composed of 6 smooth hollow residual units and 1 residual unit to learn the residual between the defogging image and the clear image; each smooth hollow residual unit is composed of 2 separable and shareable convolution layers and 2 ordinary convolution layers. The separable and shareable convolution layer divides the convolution kernel into two smaller size convolution kernels, so that the to-be-optimized weights are calculated separately, and the weights are shared to each convolution channel. In addition, the input of each smooth hollow residual unit is stacked to the output as the input of the next smooth hollow residual unit when the smooth hollow residual unit outputs, so that the network obtains high-level semantic information while retaining low-level semantic information.

[0065] The integrated gated fusion sub-network is composed of 1 gated fusion sub-network, 1 transpose convolution layer and 2 convolution layers. The gated fusion sub-network connects the outputs of the automatic encoder part and the 4th and 7th residual modules in the smooth expansion module, and takes the outputs as the outputs of the convolution layer. Finally, the convolution output is multiplied by the three part inputs and then stacked to obtain the output of the gated fusion sub-network. Then, the feature map resolution is changed to the original image resolution through 1 / 2 step upsampling of the deconvolution layer. Finally, the feature map is restored to the normal image dimension through 2 convolution layers, and the output of the haze residual is stacked with the fog image to obtain the defogging image.

[0066] The method for constructing the HF-GCANet defogging network model comprises the following steps:

[0067] The mixed hollow convolution is used to replace the smooth hollow convolution in the GCANet network model; as shown in Figure 5 The mixed hollow convolution structure schematic diagram is shown in

[0068] The above mixed hollow convolution comprises: deconvolution layer 1, deconvolution layer 2, deconvolution layer 3, deconvolution layer 4, deconvolution layer 5 and deconvolution layer 6; the input and output channel numbers of the deconvolution layer 1, the deconvolution layer 2 and the deconvolution layer 3 are all 64, the stride is 1, and the dilated rate is 1, 2 and 3 respectively; the input channel number of the deconvolution layer 4 is 64, the output channel number is 128, the stride is 2, and the dilated rate is 1; the input channel number of the deconvolution layer 5 is 128, the output channel number is 256, the stride is 2, and the dilated rate is 2; the input channel number of the deconvolution layer 6 is 256, the output channel number is 512, the stride is 2, and the dilated rate is 3. The convolution kernel size of all the hollow convolution layers is 3*3. The output of each hollow convolution layer is subjected to intra-channel normalization to accelerate the network convergence and facilitate image stylization migration.

[0069] The feature pyramid network is used to replace the integrated gated fusion sub-network in the GCANet network model. The feature pyramid network adds a top-down feature extraction structure compared with the original gated fusion sub-network. As shown inFigure 3 As shown in the figure, it is a schematic diagram of the feature pyramid network structure.

[0070] As shown in the figure, it is a schematic diagram of the HF-GCANet defogging network model structure. Figure 4 As shown in the figure, it is a schematic diagram of the HF-GCANet defogging network model structure.

[0071] 1. Input the foggy image into the convolution layer 1 to increase the number of image channels to 64;

[0072] 2. Input the 64-channel image vector into the convolution layer 2 to introduce more nonlinear parameters;

[0073] 3. Input the output of the previous step into the convolution layer 3 for down-sampling to reduce the dimension of the image vector by half;

[0074] 4. Input the image vector with reduced dimension into the HDC-FPN module to obtain a multi-scale fusion feature map;

[0075] 5. Input the multi-scale fusion feature map into the deconvolution layer 3 for up-sampling to restore the feature map to the original dimension;

[0076] 6. Input the feature map with the original dimension into the deconvolution layer 2 to introduce more nonlinear parameters;

[0077] 7. Input the output of the previous step into the deconvolution layer 1 to reduce the number of image channels to 3 and output the defogging image.

[0078] S3, based on the HF-GCANet defogging network model, combine YOLOv4 network to build GCA-YOLOv4 target detection network model;

[0079] As shown in the figure, it is a schematic diagram of the YOLOv4 network structure. Figure 7 The YOLOv4 network includes an input module, a backbone network module, a feature extraction module, and an output module. The input module includes a Mosaic unit, a CmBN unit, and a SAT unit. Data enhancement method Mosaic, cross-batch normalization, and self-adversarial training optimization method are used to enhance the input training data set. The Mosaic data enhancement method uses random scaling, random range cropping, and random position arrangement to splice four original images into one image as the data input of the input module. Random scaling and random range cropping will convert part of the large and medium-sized targets into small-sized targets, and the multi-image merging method will also increase the proportion of small targets in a single image. Therefore, data enhancement through Mosaic can effectively solve the problem of uneven distribution of different scale targets in the data set.

[0080] The backbone network module comprises a CBM unit and a CSP unit; the CBM is a minimum constituting unit in the YOLOv4 network structure, which is composed of a convolution layer, a batch normalization layer and an activation function. The activation function of the CBM unit is Leaky_Relu. The CSP unit structure is similar to the CSPNet structure, which is composed of a plurality of Res unit units, CBM units and Concat units. The Concat unit is used for the connection operation between vectors. The Res unit unit is similar to the residual structure in the ResNet and is used for increasing the network depth.

[0081] The feature extraction module comprises an SPP unit and an FPN unit; the SPP unit is similar to a spatial feature pyramid structure and is composed of three maximum pooling units with different step amplitudes, sizes and pooling sizes and a connection unit; the pooling layer pools the input features to obtain a plurality of feature vectors with different channel numbers, and finally combines all the feature vectors to obtain a fixed-size feature vector; the SPP module makes the input size of the convolution feature extraction network unrestricted, and at the same time can obtain a fixed-size feature output to match the full connection layer in the subsequent network, greatly increasing the network flexibility; the maximum pooling layers with different pooling kernel sizes can extract feature information of multiple scales, which is beneficial to improve the network detection precision. As shown in FIG. 2, it is a structural diagram of the FPN unit, the FPN module combines the top-down network with the bottom-up network of the SPP module to enhance the semantic features of the low-level and high-level. In order to strengthen the positioning feature information, the YOLOv4 increases a new bottom-up feature pyramid after the FPN module, which is combined with the top-down network in the FPN module. The PAN structure is adopted between each layer of the pyramid, and the two feature maps in the original PAN are connected and fused to obtain the next layer feature map, which is changed to connection fusion, so that the size of the feature map increases layer by layer, further improving the network feature extraction capability. Figure 6

[0082] The complete intersection over union loss function is used as the network overall loss function in the output module. The intersection over union function is shown in the following formula:

[0083]

[0084] In the formula, A represents the range of the real target frame, and B represents the range of the predicted target frame. The specific expression of the complete intersection over union loss function is shown in the following formula:

[0085]

[0086] In the formula, p represents the Euclidean distance, b represents the center point of the predicted target frame, b gt represents the center point of the real target frame, c represents the diagonal distance of the minimum circumscribed matrix between the predicted target frame and the real target frame, a represents a weighting parameter, and v represents a width-height ratio consistency parameter. The specific definition of v is shown in the following formula:​

[0087]

[0088] wherein w represents the predicted target frame width, h represents the predicted target frame height, w gt represents the real target frame width, h gt represents the real target frame height. The specific definition of a is shown in the formula:

[0089]

[0090] wherein IoU represents the intersection over union of the real target frame range and the predicted target frame range. The complete intersection over union loss function comprehensively considers the three important geometric elements of the overlapping area, the width-height ratio and the center point distance between the target frame and the predicted frame, so that the loss function regression is faster and more accurate in the case of multiple target frame overlaps and inclusions.

[0091] In the output module, YOLOv4 also uses a non-maximum suppression method based on distance intersection over union, instead of the original non-maximum suppression method. In the target detection process, the image needs to be feature-extracted through a sliding window, and after the extracted features are classified and detected, each sliding window is assigned a class score. In actual application, a large number of sliding windows will appear to contain and overlap each other, resulting in multiple detection frame results for the same target, which affects the final detection effect of the network. In order to solve the problem of false suppression of redundancy in the non-maximum suppression method, the non-maximum suppression method based on distance intersection over union uses DIOU instead of IOU as the basis for suppressing redundancy on the basis of the non-maximum suppression method, and comprehensively considers the overlap rate and center point distance between the target detection frames, so that two target detection frames with high overlap rate but far center point distance can be retained, reducing the target miss detection rate.

[0092] As shown in Figure 8 , it is a GCA-YOLOv4 target detection network model structure diagram; the GCA-YOLOv4 target detection network model comprises an input module, an HF-GCA module, a backbone network module, a feature extraction module and an output module; the HF-GCA module comprises an auto-encoder, a hybrid dilated convolution unit and a feature pyramid network unit.

[0093] S4, based on the model data set, training a GCA-YOLOv4 target detection network model;

[0094] The GCA-YOLOv4 target detection network model training process is as follows:

[0095] (1) The training set image is input into the input module. First, four images are randomly cropped and spliced into one image as training data by the Mosaic image enhancement unit. Then, the CmBN batch normalization processing is performed, and finally the weight coefficients of different positions are obtained through the SAM spatial attention mechanism.

[0096] (2) The feature image obtained by the input module is input into the HF-GCA module, and after the self-encoding, mixed hollow convolution and feature pyramid network processing, the dehazed feature image is obtained.

[0097] (3) The dehazed feature image is input into the backbone network module, and is sequentially processed by convolution and batch normalization and multiple cross-stage peer networks CSP modules to obtain the image feature vector.

[0098] (4) The feature vector is input into the feature extraction module, which is divided into three parts and sequentially processed by convolution, batch normalization, spatial pyramid pooling and channel connection to obtain a multi-scale fused feature vector.

[0099] (5) The multi-scale feature vector is input into the output, which is processed by convolution and batch normalization to obtain the prediction box position information and classification information. The prediction box position information loss function adopts the complete intersection over union loss function. The classification information loss function adopts the binary cross entropy, as shown below:

[0100]

[0101] In the formula, N represents the total number of samples, i represents the i-th sample, y i represents the binary label value of the i-th sample, and p(y i ) represents the probability of predicting belonging to y i label.

[0102] According to the loss function, the network parameters are adjusted, and the next round of training is performed back to step (1).

[0103] S5, based on the trained GCA-YOLOv4 target detection network model, the to-be-identified image is detected and classified to obtain the prediction box position information and classification information.

[0104] As Figure 10 shown, it is a GCA-YOLOv4 target detection network model target detection result schematic diagram; the method for obtaining the prediction box position information and the classification information comprises:

[0105] Step 1, input the to-be-identified image into the input module for image enhancement to obtain an enhanced image;

[0106] Step 2, input the enhanced image into the HF-GCA module to obtain a dehazed image;

[0107] Step 3, input the defogging image into the backbone network module for feature extraction to obtain a feature map;

[0108] Step 4, input the feature map into the feature extraction module for feature fusion to obtain a fusion feature vector;

[0109] Step 5, input the fusion feature vector into the output module to obtain the predicted frame position information and the classification information.

[0110] Embodiment Two

[0111] Through the above experiments on the verification set and the test set, the evaluation indexes including AP, AP50 and AP75 are used for comparative evaluation, which shows that the method proposed in the present application is superior to the mainstream target detection network SSD, CenterNet2 and YOLOv4 in terms of average precision. The specific evaluation index comparison is shown in Table 1:

[0112] Table 1

[0113]

[0114] Compared with other mainstream target detection, the present application fuses GCANet and YOLOv4 to realize low-quality underwater image fish target detection, which is of great significance to the exploration and sustainable development of marine fishery resources. The low-quality underwater image fish target detection method of the present application has strong universality and can realize more kinds of fish target detection by expanding the data set for training.

[0115] The above-described embodiments are only descriptions of the preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A low-quality underwater image fish target detection method, characterized in that, The method comprises the following steps: S1, collecting an underwater image dataset, preprocessing the image dataset to obtain a model dataset; S2, constructing a GCANet network model, and constructing an HF-GCANet defogging network model based on the GCANet network model; The GCANet network model comprises an autoencoder, a smooth expansion module, and a gate fusion subnetwork; The autoencoder is composed of three convolutional layers, the number of channels of the convolutional layers is 64, and the size of the convolution kernel is 3*3; The last convolutional layer of the autoencoder adopts 1 / 2 stride downsampling; The smooth expansion module is composed of six smooth hollow residual units and one residual unit; The gate fusion subnetwork is composed of one gate fusion subnetwork, one transpose convolutional layer, and two convolutional layers; The method for constructing the HF-GCANet defogging network model comprises: The mixed hollow convolution is used to replace the smooth hollow convolution in the GCANet network model; The feature pyramid network is used to replace the gate fusion subnetwork in the GCANet network model; The mixed hollow convolution comprises an inverse convolutional layer 1, an inverse convolutional layer 2, an inverse convolutional layer 3, an inverse convolutional layer 4, an inverse convolutional layer 5, and an inverse convolutional layer 6; The input and output channel numbers of the inverse convolutional layer 1, the inverse convolutional layer 2, and the inverse convolutional layer 3 are all 64, the stride is 1, and the dilated rate is 1, 2, and 3 respectively; The input channel number of the inverse convolutional layer 4 is 64, the output channel number is 128, the stride is 2, and the dilated rate is 1; The input channel number of the inverse convolutional layer 5 is 128, the output channel number is 256, the stride is 2, and the dilated rate is 2; The input channel number of the inverse convolutional layer 6 is 256, the output channel number is 512, the stride is 2, and the dilated rate is 3; S3, based on the HF-GCANet defogging network model, a GCA-YOLOv4 target detection network model is constructed in combination with a YOLOv4 network; The GCA-YOLOv4 target detection network model in S3 comprises an input module, an HF-GCA module, a backbone network module, a feature extraction module, and an output module; The HF-GCA module comprises an autoencoder, a mixed hollow convolution unit, and a feature pyramid network unit; S4, based on the model dataset, the GCA-YOLOv4 target detection network model is trained; S5, based on the trained GCA-YOLOv4 target detection network model, a to-be-recognized image is detected and classified to obtain prediction box position information and classification information.

2. The method according to claim 1, wherein, The method for obtaining the model dataset in S1 comprises: Collecting small fish videos in preset data, extracting each frame in the videos, converting the frames into image form, and obtaining an image dataset; Naming the image dataset according to time periods and frame numbers and converting the image dataset into YOLO format to obtain the model dataset.

3. The method of claim 1, wherein, The YOLOv4 network in S3 comprises an input module, a backbone network module, a feature extraction module, and an output module; The input module comprises a Mosaic unit, a CmBN unit and a SAT unit; The backbone network module comprises a CBM unit and a CSP unit; The feature extraction module comprises an SPP unit and an FPN unit.

4. The method of claim 1, wherein, The method for obtaining the prediction box position information and the classification information in S5 comprises: Step 1, inputting the image to be recognized into the input module for image enhancement to obtain an enhanced image; Step 2, inputting the enhanced image into the HF-GCA module to obtain a defogged image; Step 3, inputting the defogged image into the backbone network module for feature extraction to obtain a feature map; Step 4, inputting the feature map into the feature extraction module for feature fusion to obtain a fused feature vector; Step 5, inputting the fused feature vector into the output module to obtain the prediction box position information and the classification information.

Images