| The main structure of water-related infrastructure buildings is concrete structure,and the underwater structural parts of these buildings are affected by external loading factors and underwater environmental erosion during long-term service,which will cause different degrees of damage,and these damages will initially appear in the form of apparent disease.Therefore,we need to regularly inspect the apparent disease condition of underwater concrete structures.However,It is necessary to regularly conduct safety inspections on the apparent damage status of underwater concrete structures.Due to the complexity of the underwater environment and the scattering and absorption of light by water,blurring,fogging,distortion,and other phenomena can occur in the collected underwater images during the detection process,affecting the accurate identification and classification of diseases.In order to solve the problems of low disease detection and recognition rate and inaccurate classification caused by blurred,foggy,and distorted underwater image imaging in turbid waters.In this paper,we propose a deep learning-based intelligent identification method for underwater concrete structure apparent diseases,which uses image processing and target recognition techniques in deep learning to build a two-stage identification system for diseases in turbid waters.Among them,the system includes an image fusion module and a target recognition module.The disease data collected under turbid waters can be firstly used in the image fusion module to enhance the detectability of the disease,and then in the target recognition module(improved YOLO v5 s algorithm)to achieve fast and accurate localization and classification of the apparent disease of underwater concrete structures under turbid waters.For the research of this method,the main work and results of this paper are as follows:(1)Construction of underwater dataset: according to the need of experimental research and scheme comparison of the algorithm in this paper,a large number of underwater disease datasets are collected,including muddy water dataset and clear water dataset;(2)Construction of image fusion module: According to the characteristics and principles of underwater imaging,an image fusion module is constructed by using the multi-scale Retinex algorithm(MSR)and Cycle GAN algorithm in deep learning for image reconstruction and enhancement principles.To achieve the purpose of clear reconstruction of input images to improve the detectability of concrete apparent diseases.(3)YOLOv5s algorithm lightweighting process: In order to meet the need of embedding the algorithm into unmanned detection equipment in the field,the improved algorithm model with smaller number of parameters and more efficient detection.The Ghost module and Ghost bottleneck module of the lightweight network Ghostnet are introduced to improve the lightweighting of YOLOv5 s algorithm.(4)YOLOv5s algorithm accuracy improvement process: In order to further reduce the weight of the algorithm model,while improving the detection accuracy of the lightweight algorithm model.For the lightened YOLOv5 s algorithm,the Sim AM attention mechanism is improved and the most suitable combination of loss function and activation function is explored in the case of this paper’s dataset,and the best improvement scheme and the best combination of functions for the Sim AM attention mechanism are obtained through the comparative study and analysis of different combinations.Compared with the original YOLOv5 s algorithm,the final improved YOLOv5 s algorithm improves the m AP(mean Average Precision)by 4.7%,P(Precision)by 5.8%,and R(Recall)by 7.4%,while the model weights are reduced by 46.5%.(5)Test and analysis of the two-stage identification system of this paper: Ablation tests were conducted on the wisdom identification method of underwater concrete structure based on deep learning proposed in this paper to investigate the effect of the method on disease identification under turbid waters.The results showed that the m AP increased by 62.5%,the P increased by 7.7% and the R increased by 84.5% by using the deep learning-based underwater concrete structure wisdom identification method proposed in this paper compared with the test results without the method. |
