Underwater Garbage Detection Based On Deep Learning

Underwater target detection is an important part of the vision system of AUV and one of the key technologies to realize automatic underwater garbage disposal.It can identify and locate the target accurately by extracting the features from the image,and provides an important basis for subsequent tracking or capture.Traditional underwater target detection methods require artificial definition of features,which have low adaptability to target changes.In this thesis,deep learning method is adopted to research underwater garbage detection.Trash＿ICRA19 dataset and self-collected underwater garbage data are adopted to take underwater garbage and biological targets（plastic,metal,fish,etc.）as detection objects,and the research is carried out from two aspects of underwater image enhancement and detection model improvement,the main work of this thesis is as follows:Aiming at the problems of poor illumination conditions of underwater garbage images and small difference between target and background,design an adaptive brightness and contrast enhancement underwater image preprocessing algorithm,the brightness information of image background is collected by window sampling method,the saturation and brightness of image are adjusted adaptively,and use CLAHE algorithm to improve image contrast,the underwater image quality is improved effectively without affecting the real-time performance of the detection task.In order to further improve the detection accuracy of underwater garbage,a variety of methods were adopted to improve the PP-YOLO model.Firstly,an adaptive multiscale fusion structure was proposed to solve the problem that the underwater environmental target was similar to the background and difficult to locate,fully utilize the multi-scale information to improve the detection accuracy of the model,and design an adaptive channel allocation module to allocate the channel number of multi-scale fusion,an improved spatial feature filtering module is designed to suppress the inconsistencies of features at different scales,as a supplement to FPN,the detection accuracy is improved without significantly increasing the number of model parameters;Secondly,in order to further improve the positioning accuracy of the model,CIOU loss function is used to replace the common IOU loss,so that CIOU can measure the positioning loss of the target more comprehensively and effectively improve the detection ability of the model.Finally,experiments were carried out to verify the effectiveness of each part of the improved model.The experimental results show that the improvement of each part can effectively improve the accuracy,the detection accuracy of the improved model reaches92.1% on trash＿ICRA19 dataset,in self-collected data,the detection accuracy reached89.5%,and the detection speed is 25 FPS,which meets the requirements of practical application.