Research On Real-time Target Detection And Recognition Technology For Shallow Underwater Biology Based On Deep Learning

China is rich in underwater fishery resources,including sea urchins,shellfish and sea cucumbers and other aquatic animals.These resources are of great significance for ensuring national food security and maintaining marine ecological environment.Hightech underwater survey equipment can be used to strengthen the survey of underwater fishery resources for scientific development and utilization.Among them,the Remote Operated Vehicle(ROV)is used to shoot underwater images with a high-resolution camera,and then the fishery resources in the enhanced images are detected and identified by image enhancement algorithm and target detection algorithm based on deep learning.Aiming at the serious color distortion and rough contour of underwater optical images,and for the purpose of real-time,autonomous and high accuracy detection and identification of underwater organisms,this thesis proposes a depth enhancement algorithm for underwater images in shallow water and an improved algorithm for underwater biological detection and identification based on YOLOv5 s,and develops a ROV with a high-definition camera,which can realize underwater image shooting and transmit to ground workstation for real-time detection.The main contents of this study are as follows:(1)In view of the inherent problems of underwater optical image,such as color distortion and rough contour,this thesis presents a real-time image enhancement method for underwater optical image,which realizes real-time enhancement of underwater image quality and greatly improves the quality of underwater image.The information entropy,UCIQE and UIQM are improved by 2.4256,0.2268 and 0.8258 respectively compared with the original image,and are higher than other comparison algorithms.Aiming at the fact that the existing open underwater biological dataset can not provide high quality data service,this study collects and collects 1359 underwater biological images on the Internet,and fuses them with two open dataset with uneven proportion of samples to obtain 10570 underwater biological images.Using affine transform and Mosaic image data enhancement method,the data set is enhanced effectively,which lays a foundation for the training and validity verification of the model.(2)In order to solve the problems of low accuracy in detecting and identifying underwater biological targets by using YOLOv5 s which can meet the real-time requirements directly due to unclear color and outline of underwater optical images,an underwater biological target detection method is proposed by improving YOLOv5 s.Firstly,the fast spatial pyramid(FSPP)module is introduced to improve the model’s ability to describe texture features;secondly,the Double Swin Transformer(DST)module is introduced into the network to improve the model’s adaptability to changes in object shape and size;finally,the neck structure of the network is adjusted to improve the model’s ability to detect small targets in images.These improved methods effectively improve the ability of the improved YOLO v5 s to detect and identify underwater biological targets in shallow water,with an average accuracy of 85.7%.(3)Aiming at the problems of low efficiency,high cost and high danger of traditional underwater biological detection methods,this task team independently developed a sixdegree-of-freedom ROV.The core hardware of the robot is raspberry pie and STM32F407 small system board.Through the Raspberry Pie,the images taken by the HD camera installed on the robot end can be remotely captured by the ground station through the LAN video push technology.