Underwater structure apparent disease identification method and system based on deep learning

Abstract

The application provides a kind of underwater structure apparent disease identification method and system based on deep learning, wherein, method includes: according to the combination of fusion model and target recognition model, generate the pre-set underwater structure apparent disease identification model;Underwater apparent disease is identified by the pre-set underwater structure apparent disease identification model;The fusion model is built by improved CycleGAN model and multi-scale Retinex algorithm (MSR) network, for converting underwater image into clear and clear image with obvious features;The target recognition model is obtained by YOLOv5 model, for realizing the positioning and classification of underwater structure apparent disease.The present application solves the problems of inaccurate classification and low recognition accuracy of underwater structure apparent disease by the underwater structure apparent disease identification model, caused by factors such as camera imaging blur, insufficient contrast, dispersion and noise in complex water area.

Description

TECHNICAL FIELD

[0001] The present document relates to the technical field of underwater structure apparent disease identification, and particularly relates to an underwater structure apparent disease identification method and system based on deep learning. BACKGROUND

[0002] In recent years, with the development of economy and technology, a large number of water-related infrastructures such as bridges, dams and wharfs have been built in China. During the service process, these structures are subjected to load and environmental erosion for a long time, which may cause damage of different degrees, affecting the safe operation and service life of the structures, and even causing major accidents such as bridge collapse. Therefore, in order to ensure the safety of the structures, it is necessary to regularly detect the diseases of the underwater foundations. Traditionally, the apparent diseases of underwater structures are determined by manual exploration, which not only has personal safety risks, but also can only roughly estimate the damage of the structure surface. Meanwhile, during the data collection process by using underwater cameras and other devices, the detection results are easily affected by factors such as water turbidity, resulting in a significant decrease in disease identification accuracy. In recent years, with the rapid development of computer vision and artificial intelligence technology, more and more intelligent algorithms have been applied to structure disease detection. Although many domestic and foreign experts and scholars have made a lot of researches and achieved certain results, the effect is not ideal in the actual engineering application. Therefore, it is of great engineering and economic significance to effectively apply artificial intelligence technology to underwater structure detection and realize efficient and accurate detection of structure apparent diseases in complex underwater environment.

[0003] In recent years, deep learning networks have made significant progress in image processing and target recognition classification, but there are still some limitations, and there are few models suitable for complex water areas. In underwater environment, due to the scattering and absorption of water medium and the influence of water flow, underwater images are blurred, fogged, distorted and have low resolution, which makes it difficult to carry out underwater structure health detection. When applying target recognition algorithm to underwater structure disease detection, due to the influence of imaging equipment, the recognition rate is not high and the classification is not accurate in complex water areas. The existing underwater image enhancement processing algorithm is based on a specific environment to construct a processing model, and there are problems of relying on degradation model and low efficiency, which limits the application range and real-time performance. Therefore, it is necessary and urgent to find a fast and effective method to process underwater images and obtain images with high signal-to-noise ratio and good clarity. SUMMARY

[0004] One or more embodiments of the present specification provide an underwater structure apparent disease identification method and system based on deep learning, aiming to solve the above problems.

[0005] The embodiment of the present application provides a kind of underwater structure apparent disease identification method based on deep learning, comprising:

[0006] S1, according to the combination of preset fusion model and preset target recognition model, generate the preset underwater structure apparent disease identification model;

[0007] S2, the underwater apparent disease is identified by the preset underwater structure apparent disease identification model;

[0008] Wherein, the preset fusion model is built by improved CycleGAN model and multi-scale Retinex algorithm network, for converting underwater image into feature obvious and clear image;The preset target recognition model is obtained by YOLOv5 model, for realizing the positioning and classification of underwater structure apparent disease.

[0009] The embodiment of the present application provides a kind of underwater structure apparent disease identification system based on deep learning, comprising:

[0010] Model building module, for generating the preset underwater structure apparent disease identification model according to the combination of preset fusion model and preset target recognition model;

[0011] Identification module, for identifying the underwater apparent disease by the preset underwater structure apparent disease identification model;

[0012] Wherein, the preset fusion model is built by improved CycleGAN model and multi-scale Retinex algorithm network, for converting underwater image into feature obvious and clear image;The preset target recognition model is obtained by YOLOv5 model, for realizing the positioning and classification of underwater structure apparent disease.

[0013] With the embodiment of the present application, the image data collected from complex water area is first enhanced by image processing network to enhance the detectability of disease, and the weight fusion of CycleGAN enhanced original underwater image and multi-scale Retinex algorithm restored object essential feature image is realized in stage one;Again through YOLOv5 target recognition algorithm, the purpose of rapid and accurate positioning and classification of underwater structure apparent disease type under complex water area is achieved.The present application breaks the limitation of traditional method, provides support and help for later deep learning algorithm carried on underwater equipment for underwater structure detection work, is a major innovation and breakthrough to present underwater structure disease detection method. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the one or more embodiments of the present specification or the prior art, the drawings needed to be used in the embodiment or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present specification, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0015] Figure 1 Flow chart of the underwater structure apparent disease identification method based on deep learning of the embodiment of the present application;

[0016] Figure 2 Schematic diagram of the two-stage intelligent identification system of the apparent disease of the underwater structure of the embodiment of the present application

[0017] Figure 3 Flow chart of the two-stage disease intelligent identification system of the embodiment of the present application;

[0018] Figure 4 Display of the three common apparent disease types of the underwater structure of the embodiment of the present application;

[0019] Figure 5 CycleGAN training process of the embodiment of the present application;

[0020] Figure 6 Image processing fusion process of the embodiment of the present application;

[0021] Figure 7 Retinex model schematic diagram of the embodiment of the present application;

[0022] Figure 8 Improved generator network schematic diagram of the embodiment of the present application;

[0023] Figure 9 Unimproved generator network schematic diagram of the embodiment of the present application;

[0024] Figure 10 Schematic diagram of the residual structure in the generation network of the embodiment of the present application being replaced by an attention mechanism module;

[0025] Figure 11 Schematic diagram of the underwater structure apparent disease identification system based on deep learning of the embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to make the person skilled in the art better understand the technical solutions in one or more embodiments of the present specification, the technical solutions in one or more embodiments of the present specification will be described clearly and completely in the following with reference to the drawings in one or more embodiments of the present specification. Obviously, the described embodiments are only a part of the embodiments of the present specification, not all. Based on one or more embodiments of the present specification, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the present document.

[0027] Method embodiment one

[0028] The embodiment of the present application provides a kind of underwater structure apparent disease identification method based on deep learning, Figure 1 It is the flow chart of the underwater structure apparent disease identification method based on deep learning of the embodiment of the present application, according to Figure 1 The underwater structure apparent disease identification method based on deep learning of the embodiment of the present application specifically includes:

[0029] S1, according to the combination of pre-set fusion model and pre-set target identification model, generate pre-set underwater structure apparent disease identification model;

[0030] S2, the underwater apparent disease is identified by pre-set underwater structure apparent disease identification model;

[0031] Among them, the pre-set fusion model is built by improved CycleGAN model and multi-scale Retinex algorithm network, for converting underwater image into feature obvious and clear image;Pre-set target identification model is obtained by YOLOv5 model, for realizing the positioning and classification of underwater structure apparent disease.

[0032] The construction step of pre-set underwater structure apparent disease identification model in step S1 in the embodiment of the present application specifically includes:

[0033] S21, image data of different underwater environment is collected, image data is divided into clear water set and muddy water set according to clarity, clear water set is divided into clear water training set and clear water verification set according to predetermined proportion, and muddy water set is divided into muddy water training set and muddy water test set;In the embodiment of the present application, data set is divided into A (clear water) and B (muddy water) set according to turbidity, then A set is divided into training set A1 and verification set A2 according to certain proportion, and B set is divided into training set B1 and test set B2.

[0034] S22, train the improved CycleGAN model through the clear water training set and the muddy water training set, obtain the trained CycleGAN generator; build an image generation fusion model by combining the trained CycleGAN generator with the multi-scale Retinex algorithm network; in this stage, the clear water image set and the muddy water image set are used for training, as shown in Figure 2 The number of pictures of the training set A1 and the training set B1 is set to be 1:1, the CycleGAN model of the first stage is trained in a generative adversarial manner, and a generator (Generator Network) capable of realizing high-definition restoration of underwater structure disease pictures is obtained; the trained CycleGAN generator network is combined with the multi-scale Retinex algorithm network (MSR) to build an image generation fusion model.

[0035] S23, train the YOLOv5 model through the clear water set, and optimize the YOLOv5 model weight, so as to obtain a target recognition model capable of realizing high-precision classification of underwater structure diseases; in this stage, only the clear water image set is used for training, the clear water data set A (the training set A1 and the verification set A2, the number of pictures of A1 and A2 is 8:2) is used to train the YOLOv5 model of the second stage, the model weight is optimized, and a target recognition model capable of realizing high-precision classification of underwater structure diseases is obtained;

[0036] S24, the fusion model and the target recognition model are connected in series to form an underwater structure apparent disease recognition model. Specifically, the constructed image processing network structure model and the trained model weight of the YOLOv5 are connected in series to form the entire underwater intelligent recognition system;

[0037] After the system is connected in series, the turbid water test set B2 is used as the recognition system test set, which is first subjected to image generation fusion through the image processing network structure model, and then subjected to two-stage automatic recognition through the Yolov5 model.

[0038] The underwater structure apparent disease recognition model of the embodiment of the application is realized through two stages, Figure 2 The underwater structure apparent disease two-stage intelligent recognition system of the embodiment of the application is shown in FIG. 1, according to Figure 2 It can be seen that the underwater structure apparent disease two-stage intelligent recognition system combining the image processing network and the target recognition network includes two stages:

[0039] Stage one: the image data collected under complex water areas is respectively subjected to clear image generation through the improved CycleGAN algorithm and underwater object essential feature image generation through the multi-scale Retinex algorithm, and then the images generated by the two algorithms are fused by weight to form a feature obvious and clear image, so as to enhance the detectability of the disease;

[0040] Stage two: through the YOLOv5 model to realize the rapid, accurate positioning and classification of the apparent disease type of the underwater structure in complex water area.

[0041] The process of the embodiment of the application will be described below in conjunction with specific examples. As shown in the figure, it is a flow chart of the two-stage disease intelligent identification system of the embodiment of the application: Figure 3

[0042] I. Establishing a database

[0043] 1. Data collection and production:

[0044] The data set is produced by simulating the three typical disease characteristics (cracks, exposed reinforcement and spalling) of the underwater structure under the action of the real environment (turbid water area and clear water area) in the laboratory. The test specimen comes from the beam body damaged after the ballast test, which is placed in the pool after being treated for a period of time, and the underwater camera is used to collect disease data. According to the apparent characteristic category, the data set is produced and named as: crack (L), spalling (B) and exposed reinforcement (G) to facilitate the subsequent classification of the apparent disease of the underwater structure.

[0045] The three types of common apparent diseases of the underwater structure are shown in the figure: Figure 4

[0046] Cracks: transverse cracks, longitudinal cracks, diagonal cracks or network cracks on the surface of the underwater structure

[0047] Exposed reinforcement: exposed reinforcement and reinforcement corrosion of the underwater structure

[0048] 2. Divide the data set:

[0049] The underwater disease image is analyzed, and the image data set under the two kinds of water environment is divided; the clear water area data set (A) is divided into training set A1 and verification set A2 according to the ratio of 8:2; the turbid water area data set (B) is divided into training set B1 and test set B2 according to the ratio of 8:2, and the data division of each type is as shown in Table 1:

[0050] Table 1: Disease statistics of each type of data

[0051]

[0052] The data is divided into training set and test set, which aims to hope that the model has good ability to predict new data; the data is divided into A set and B set, which aims to train the generator of the CycleGAN network model, so that the model has the ability to clearly process the images under turbid water area, and the intelligent recognition network built can be suitable for different turbid water environments.

[0053] ​​II. Establishment of two-stage intelligent detection system

[0054] The proposed two-stage intelligent recognition system of underwater structure apparent disease based on deep learning algorithm consists of two stages: image generation fusion based on deep learning image processing part (stage one) and automatic disease recognition detection based on deep learning target recognition aspect (stage two). Once we obtain the original disease image from the surface of the underwater target, the network model generates the corresponding clear image at the first time; then, the apparent disease of the generated clear image is identified through the automatic detection of the target recognition aspect, so as to achieve the detectability of the apparent disease recognition.

[0055] The image processing (stage one) and the target recognition (stage two) of the system are two independent training processes.

[0056] 1. Stage one training

[0057] Firstly, in this research, the data set used for stage 1 training is training set A1 (2400 clear water pictures) and training set B1 (2400 muddy water pictures). The improved CycleGAN trains the style difference between the muddy images and the clear images, so that the generator (Generator Network) can well convert the blurred pictures into clear pictures. The generation and confrontation training process is as follows Figure 5 A , D B represents the discriminator of CycleGAN, and Generator Network represents the generator of CycleGAN, and the training is the generation and confrontation process of the generator and the discriminator.

[0058] The embodiment of the application proposes an improved CycleGAN model, in order to make the network model of stage 1 better adapt to the image conversion of underwater environment, the generator structure of the original CycleGAN is optimized, and the network diagram after optimization is as follows Figure 8 , Figure 9 The improved generator network diagram is shown in the figure, wherein the main improvements have three parts, as follows:

[0059] (1) The CycleGAN architecture for single image deblurring is enhanced by adding cycle perception consistency loss in addition to cycle consistency loss. The main idea of this loss is to compare images in feature space rather than pixel space. The original image and the reconstructed cycle image are compared in two spaces, wherein the cycle consistency loss ensures a high PSNR value, and the perception loss maintains the clarity of the image.

[0060] ​(2) SiLU activation function has the characteristics of no upper bound, lower bound, smoothness and non-monotonicity, which all have a positive effect on the performance of similar activation functions. The effect is better than ReLU, especially in deep networks. Therefore, the ReLU activation function in the original generator is replaced by SiLU.

[0061] (3) Replace the residual structure in the generator network with the attention mechanism module (CBAM), which is a simple and effective attention module for feedforward convolutional neural networks. Given an intermediate feature map, our module will infer attention maps along two independent dimensions (channels and space) in turn, and then multiply the attention maps with the input feature map for adaptive feature modification. Finally, the texture of the image reconstructed by the model is clearer and more specific. The comparison before and after the replacement is shown in Figure 10 .

[0062] 2. Construction of phase one network

[0063] In this part of the image processing network, because the CycleGAN algorithm loses part of the feature information in the process of restoring underwater images, the restored image is fused with the object essence feature image generated by the multi-scale Retinex algorithm (MSR), making the final generated image features clearer and more obvious. To achieve this process, we build the trained CycleGAN generator network and the multi-scale Retinex algorithm network, and then use the images generated by the two algorithms to fuse the two images according to a certain weight ratio by using the functions provided in OpenCV. The image processing network structure model of phase one is constructed, so that the fused underwater image will meet the high-precision detection requirements. The image processing fusion process of the model is shown in Figure 6 .

[0064] Retinex is a commonly used image enhancement method based on scientific experiments and scientific analysis. The basic idea of Retinex theory is that the intensity of light determines the dynamic range of all pixel points in the original image, and the inherent properties of the original image are determined by the reflection coefficient of the object itself, that is, the product of the reflection image and the light image is the original image. Therefore, the idea of Retinex is to remove the influence of light and retain the inherent properties of the object. Its core principle is as follows Figure 7 , assuming that the image observed by the observer is l(x, y), the expression is:

[0065] I(x, y) = L(x, y) x R(x, y)

[0066] Where L(x, y) represents the incident light image, which directly determines the dynamic range that the pixels in the image can reach, R(x, y) represents the reflection property image of the object, that is, the intrinsic property of the image, and l(x, y) represents the image that the human eye can receive.

[0067] Multi-scale Retinex algorithm (MSR) is a multi-scale Retinex algorithm proposed by Jobson and other researchers to overcome the shortcomings of single-scale Retinex model (SSR). The algorithm combines the results of different scales by linear combination, fully considering local information and overall information. The main idea of this algorithm is to estimate the illumination component by weighted average of several different scale center-surround functions. MSR algorithm can produce a single output image with good dynamic range compression, color stability and good tone recovery. The formula of MSR algorithm is:

[0068]

[0069] Where N represents the number of scales, N = 1 for SSR, and to ensure that the MSR algorithm has the advantages of high and low scales, the value of N is generally selected as 3, and the effect is better when three different scale Gaussian filters are used for filtering the original image. k is the weighting coefficient when the kth scale is weighted, which needs to satisfy W1 = W2 = W3 = 1 / 3 is usually taken, and it is found through experiments that when taking the average, it can be applied to a large number of low-illumination images, and the operation is simple.

[0070] 3. Stage two training

[0071] The second stage YOLOv5 network is trained and verified by the clear water data set (3000 images) collected in the laboratory. As shown in Table 1, the training set A1 (2400 images) and the verification set A2 (600 images) are divided in the ratio of 8:2. The training set A1 and the verification set A2 are used to train the YOLOv5 model for 150 rounds, so that the YOLOv5 model can well identify and classify various underwater structural appearance diseases.

[0072] 4. Series connection of two-stage system

[0073] Then a complete underwater appearance disease identification system is built by combining the constructed image processing network structure model with the YOLOv5 model, which is applied to actual underwater detection.

[0074] Three, disease identification

[0075] The test set B2 (600 images) is input into the built two-stage intelligent recognition system of underwater structure apparent disease to realize the recognition and classification of image data of the underwater structure apparent disease. The recognition result is as follows Figure 11 As can be seen from the figure, the two-stage intelligent recognition system of underwater structure apparent disease based on the deep learning algorithm can well reconstruct the image under turbid water and accurately classify it.

[0076] Test result mAP 0.5 The value (indicating the mAP value when the IOU threshold is 0.5), and the precision (Precision) are shown in Table 2.

[0077] Table 2 result index comparison

[0078]

[0079] As can be seen from the results in the table, the embodiment of the present application can achieve more than 85% in mAP 0.5 value and precision when detecting images under turbid water, and can effectively recognize and classify the images under turbid water. At the same time, the size of the improved CycleGAN model weight is 17.2% of the original CycleGAN. This greatly reduces the requirement of the underwater intelligent recognition model on computer hardware, and can better be loaded on underwater equipment for underwater structure apparent disease detection.

[0080] By adopting the embodiment of the present application, the following beneficial effects are achieved:

[0081] By adopting the embodiment of the present application, the image data collected from complex underwater areas is first enhanced by the image processing network to improve the detectability of the disease. In the first stage, the weight fusion of the original underwater image enhanced by CycleGAN and the object essential feature image restored by the multi-scale Retinex algorithm is realized. Then, the YOLOv5 target recognition algorithm is used to achieve the purpose of rapid and accurate positioning and classification of the underwater structure apparent disease type under complex underwater areas. The present application breaks the limitations of traditional methods, provides support and help for loading deep learning algorithm on underwater equipment for underwater structure detection in the later stage, and is a major innovation and breakthrough in the field of underwater structure disease detection method.

[0082] System embodiment

[0083] The embodiment of the present application provides a deep learning-based underwater structure apparent disease recognition model construction system, Figure 11 The embodiment of the present application provides a deep learning-based underwater structure apparent disease recognition model construction system, Figure 11As shown, the deep learning-based underwater structure apparent disease identification model construction system of the embodiment of the present application specifically comprises:

[0084] The model construction module 110 is configured to combine the preset fusion model with the preset target identification model to generate a preset underwater structure apparent disease identification model.

[0085] The identification module 111 is configured to identify underwater apparent diseases through the preset underwater structure apparent disease identification model.

[0086] The preset fusion model is built by an improved CycleGAN model and a multi-scale Retinex algorithm network, and is used to convert underwater images into images with obvious and clear features; the preset target identification model is obtained by a YOLOv5 model, and is used to realize positioning and classification of underwater structure apparent diseases.

[0087] The model construction module 110 of the embodiment of the present application specifically comprises:

[0088] The experimental data acquisition submodule is configured to collect image data of different underwater environments, divide the image data into a clear water set and a turbid water set according to the clarity, divide the clear water set into a clear water training set and a clear water validation set according to a predetermined ratio, and divide the turbid water set into a turbid water training set and a turbid water test set.

[0089] The generator submodule is configured to train the improved CycleGAN model through the clear water training set and the turbid water training set, and obtain a trained CycleGAN generator.

[0090] The fusion submodule is configured to build an image generation fusion model by the trained CycleGAN generator and the multi-scale Retinex algorithm network.

[0091] The identification submodule is configured to train the YOLOv5 model through the clear water set, optimize the YOLOv5 model weight, and obtain a target identification model capable of realizing high-precision classification of underwater structure diseases.

[0092] The model generation submodule is configured to connect the fusion model and the target identification model in series to form an underwater structure apparent disease identification model.

[0093] The experimental data acquisition submodule specifically comprises:

[0094] The image data of turbid water areas are acquired as the turbid water set, the image data of clear water areas are acquired as the clear water set, and the turbid water set and the clear water set are marked according to underwater apparent diseases.

[0095] The marked turbid water set and the clear water set are divided, the clear water set is divided into a clear water training set and a clear water verification set according to a ratio of 8:2, and the turbid water set is divided into a turbid water training set and a turbid water test set according to a ratio of 8:2.

[0096] The generator sub-model specifically comprises:

[0097] The CycleGAN model is improved, specifically including: adding a cycle perception consistency loss to the cycle consistency loss of the CycleGAN model, for enhancing the CycleGAN architecture of single image deblurring; replacing the ReLU activation function in the CycleGAN model with a SiLu activation function; replacing the residual structure in the CycleGAN model with an attention mechanism module;

[0098] The improved CycleGAN model is trained through the clear water training set and the turbid water training set, so as to obtain a trained improved CycleGAN model, and the trained improved CycleGAN model is used to convert a blurred picture into a clear picture.

[0099] The fusion sub-model specifically comprises:

[0100] The multi-scale Retinex algorithm network is used to generate an object essential feature image;

[0101] The object essential feature image of the clear picture converted by the CycleGAN model is fused according to a specific weight ratio.

[0102] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for identifying apparent defects of underwater structures based on deep learning, characterized in that, The application relates to an underwater structure apparent disease identification method. S1, a preset underwater structure apparent disease identification model is generated by combining a preset fusion model and a preset target identification model; S2, the underwater apparent disease is identified through the preset underwater structure apparent disease identification model; The preset fusion model is built by combining an improved CycleGAN model and a multi-scale Retinex algorithm network, and is used for converting an underwater image into a clear image with obvious features; the preset target identification model is obtained through a YOLOv5 model, and is used for realizing positioning and classification of the underwater structure apparent disease. The construction steps of the preset underwater structure apparent disease identification model in S1 specifically include: S21, image data of different underwater environments is collected, the image data is divided into a clear water set and a turbid water set according to the definition, the clear water set is divided into a clear water training set and a clear water verification set according to a predetermined proportion, and the turbid water set is divided into a turbid water training set and a turbid water test set; S22, the improved CycleGAN model is trained through the clear water training set and the turbid water training set, and a trained CycleGAN generator is obtained; the trained CycleGAN generator is combined with the multi-scale Retinex algorithm network to build an image generation fusion model; S23, the YOLOv5 model is trained through the clear water set, and the YOLOv5 model weight is optimized to obtain a target identification model; S24, the fusion model and the target identification model are connected in series to form the underwater structure apparent disease identification model; The training of the improved CycleGAN model through the clear water training set and the turbid water training set in S22 specifically includes: The CycleGAN model is improved, specifically including: adding a cycle perception consistency loss to the cycle consistency loss of the CycleGAN model, which is used for enhancing the CycleGAN architecture of single-image deblurring; replacing the ReLU activation function in the CycleGAN model with a SiLu activation function; and replacing the residual structure in the CycleGAN model with an attention mechanism module; The improved CycleGAN model is trained through the clear water training set and the turbid water training set, so that a trained CycleGAN generator is obtained, and the trained CycleGAN generator is used for converting a blurred picture into a clear picture; The training of the improved CycleGAN model through the clear water training set and the turbid water training set in S22 specifically includes: An object essential feature image is generated through the multi-scale Retinex algorithm network; The clear picture converted by the trained CycleGAN generator and the object essential feature image are fused according to a preset weight ratio.

2. The method of claim 1, wherein, The S21 specifically includes: Image data of turbid water areas is obtained as the turbid water set, image data of clear water areas is obtained as the clear water set, and the turbid water set and the clear water set are marked according to underwater apparent diseases. The marked muddy water set and the clear water set are divided, and the clear water set is divided into a clear water training set and a clear water verification set according to a ratio of 8:2, and the muddy water set is divided into a muddy water training set and a muddy water test set according to a ratio of 8:

2. 3.A system for constructing a deep learning-based underwater structure apparent disease identification model, characterized by, Comprise: The model construction module is used for generating a preset underwater structure apparent disease identification model by combining a preset fusion model with a preset target identification model; The identification module is used for identifying underwater apparent diseases through the preset underwater structure apparent disease identification model; The preset fusion model is built by an improved CycleGAN model and a multi-scale Retinex algorithm network, and is used for converting underwater images into images with obvious and clear features; the preset target identification model is obtained by a YOLOv5 model, and is used for realizing positioning and classification of underwater structure apparent diseases; The model construction module specifically comprises: The experimental data acquisition submodule is used for collecting image data of different underwater environments, dividing the image data into a clear water set and a muddy water set according to the clarity, dividing the clear water set into a clear water training set and a clear water verification set according to a predetermined ratio, and dividing the muddy water set into a muddy water training set and a muddy water test set; The generator submodel is used for training the improved CycleGAN model through the clear water training set and the muddy water training set, and obtaining a trained CycleGAN generator; The fusion submodel is used for building an image generation fusion model by the trained CycleGAN generator and the multi-scale Retinex algorithm network; The identification submodule is used for training the YOLOv5 model through the clear water set, optimizing the YOLOv5 model weight, and obtaining a target identification model; The model generation submodule is used for connecting the fusion model and the target identification model in series to form an underwater structure apparent disease identification model; The generator submodel specifically comprises: The CycleGAN model is improved, specifically including: adding a cycle perception consistency loss outside the cycle consistency loss of the CycleGAN model, which is used to enhance the CycleGAN architecture of single image deblurring; replacing the ReLU activation function in the CycleGAN model with the SiLu activation function; replacing the residual structure in the CycleGAN model with an attention mechanism module; The improved CycleGAN model is trained through the clear water training set and the muddy water training set, so as to obtain a trained CycleGAN generator, which is used for converting blurred pictures into clear pictures; The fusion submodel specifically comprises: An object essence feature image is generated through the multi-scale Retinex algorithm network; The clear picture converted by the trained CycleGAN generator and the object essence feature image are fused according to a preset weight ratio.

4. The system of claim 3, wherein, The experimental data acquisition submodule specifically comprises: Image data of turbid water areas are obtained as the muddy water set, and image data of clear water areas are obtained as the clear water set, and the muddy water set and the clear water set are marked according to underwater apparent diseases; The marked turbid water set and the clear water set are divided, the clear water set is divided into a clear water training set and a clear water verification set according to a ratio of 8:2, and the turbid water set is divided into a turbid water training set and a turbid water test set according to a ratio of 8:2.

Images