Research On Underwater Image Enhancement And Underwater Optical Target Detection Based On Deep Learning

The excessive consumption of land resources makes people’s exploration of marine resources imminent,but the exploration of the underwater environment is highly risky and unknown.There will be great danger to life if the underwater operations are completed directly through manual underwater exploration.Therefore,the development of underwater robots is of great significance for underwater exploration.Whether the underwater robot can intelligently perform underwater operations needs to be solved first is the underwater robot’s vision system module.Therefore,this paper studies how to improve the performance of the underwater robot vision system module.Underwater image enhancement technology and underwater target detection technology are important components of underwater robot vision technology,the main research contents of this paper are as follows:In terms of underwater image enhancement,the design principles of a variety of underwater image enhancement algorithms with better enhancement effects are analyzed in detail,and the algorithm design is classified according to whether it is suitable for underwater imaging.Through code simulation,compare and show the enhanced performance of each algorithm.The fusion deep learning model designed a lightweight convolutional neural network suitable for underwater image enhancement,and combined the public underwater data set and Enhancement of Underwater Images with Statistical Model algorithm to build an underwater data set for the network training.Through subjective and objective analysis,it is proved that the proposed underwater image enhancement algorithm has superiority in enhancement effect,scene applicability and operating efficiency.In terms of underwater target detection,for the four types of detected targets(echinus,holothurian,scallops and starfish)in the public data set provided by the National Natural Science Foundation of China Underwater Robot Competition,Labelme labeling software is used to label the true values.The YOLO V5 algorithm is improved in two aspects,in terms of data set,the proposed deep learning-based underwater image enhancement algorithm module is firstly added,which effectively enriches the scene types of the training data set and greatly increases the number of detected objects in the data set,making the network training more robust.Secondly,the Mosaic algorithm module is added.Through image splicing and scaling,the number of small targets in the data set is effectively increased,and the detection ability of the algorithm in small target scenes is enhanced.In terms of network,the Convolutional Block Attention Module attention mechanism module is added,so that the attention of the early feature extraction of the network is more allocated to the detected target.By comparing the detection results before and after each improvement,the effectiveness of the algorithm improvement is proved.Through the comparative analysis before and after each improvement,the effectiveness of the improved algorithm is verified.The improved algorithm is tested and compared with the single-stage object detection algorithm and the two-stage object detection algorithm.It shows that the improved underwater optical target detection algorithm has high accuracy while still ensuring real-time performance,which can meet the real-time detection tasks of underwater robots in most scenarios.