Research On Underwater Target Detection Algorithm Based On Deep Learning

The explosive growth of data from underwater optical images,as a result of the advancement of underwater robot technology,has sparked an ever-growing research heat.These images can be of great assistance to many underwater operations,such as inspecting and repairing artificial underwater facilities with live video feeds from underwater robots,biologists can study Marine life using underwater images,and geologists can use images of the ocean floor to detect resources.The most frequently employed auxiliary tools in these tasks are the underwater image processing and target detection algorithms.However,due to the special underwater environment,the image can suffer considerable deterioration.The traditional image enhancement algorithm is hard to use for intricate underwater images,and the current mainstream target detection model has more issues with error and miss detection when applied to underwater images.This thesis delves into the realm of underwater image enhancement and target detection algorithms,based on the principles and techniques of deep learning.The research accomplished is as follows:In light of the limited efficacy of underwater image enhancement algorithms and their scarcity,this thesis proposes an amalgamation of two-color space restoration and multi-level decoding structure as a means of enhancing underwater images.The image quality is improved from color and contrast.This thesis puts forward a multi-scale feature learning-based underwater target detection algorithm,with the aim of resolving the issue of fuzzy targets,overlapping targets,and small targets being overlooked and misidentified by existing algorithms.To begin,the dual backbone network is used to extract the depth and texture features,which are then processed together.Through the coordinate attention mechanism,the location information can get better attention,so as to detect the target more accurately.In addition,this algorithm increases the number of prediction heads to 4,which can better detect small targets and reduce the resource occupation of the network model,so that the model is suitable for embedded devices with limited hardware conditions.Compared with other advanced target detectors,it further improves the detection accuracy of underwater targets and can effectively detect targets with good detection accuracy and real-time performance.This thesis proposes an algorithm for improving underwater image quality,which experiments demonstrate can be effectively implemented.The proposed underwater target detection algorithm has a higher Mean Average Precision(m AP)score than YOLOv5 and other comparison algorithms.