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

In recent years,underwater target detection technology has developed rapidly and has been widely used in various fields.Image vision systems make up for the shortcomings of acoustic vision due to their advantages,such as the ability to perceive at close range and the richer image information obtained,and have become an important branch of underwater target detection studied by experts and scholars at home and abroad in recent years.Underwater environments are often affected by the scattering and refraction of light by the water medium,absorption effects,and underwater suspensions,which often result in blurred details,low contrast,and unclear edges,the performance of existing underwater target detection methods is limited by environmental background factors and underwater image degradation.To address these issues,this thesis focuses on underwater image enhancement techniques based on deep learning and underwater target detection techniques based on deep learning to solve some of the existing problems in the fields of underwater image enhancement and underwater target detection.The main work in this thesis is as follows:Firstly,the research background and significance of underwater image enhancement and underwater target detection technologies are explained,the current status of domestic and international research on underwater image enhancement and underwater target detection technologies based on traditional methods and deep learning methods is studied in depth,and the advantages and shortcomings of various underwater image enhancement and underwater target detection technologies are analyzed.Then,a method of underwater image enhancement based on an improved generative adversarial network is proposed for the image degradation problem that can exist in the images acquired by direct underwater photography.This method replaces part of the activation function of the generated network module with the SeLU activation function to provide more abundant features for the feature map;replaces the feature extraction network with the RepVGG network to speed up the reasoning of the model;introduces L1 loss of the gradient image to enrich the edge information and structure information of the generated image;and improves the restoration effect of underwater degraded images.It is proved through experiments that the underwater degraded images recovered by the underwater image enhancement method proposed in this thesis can achieve high subjective and objective evaluation criteria,with peak signal to noise ratio and structural similarity reaching 28.90 and 0.91,respectively,which are 26.9% and 39.1% higher than those of FUnIE-GAN and Shallow-UWnet,respectively,and the inference speed of the improved model is 14.5 milliseconds,which is 48.3% of the pre-improvement model and only 6.7% of FFA-Net.The clarity of the restored image has increased by 14.9% compared to the original image.Finally,an underwater target detection method based on improved Faster R-CNN is presented to solve the complex environment of underwater background.This method replaces the feature extraction network with the Res Net-50 network to eliminate the problem of rapid decline in recognition accuracy caused by the increase in network layers;introduces the CBAM Attention Mechanism so that the model can reduce the interference of factors such as the background and use the limited computational resources to obtain more effective information and improve the target detection accuracy;and replaces the original border loss function with the CIo U border loss function to further improve the accuracy of the target detection frame.It is demonstrated through experimental results that the underwater target detection method based on the improved Faster R-CNN proposed in this thesis can detect underwater targets more accurately.On the National Underwater Robotics Competition dataset,the mAP@.5:.95 was improved by 1.55% compared with the Faster R-CNN algorithm,and there was more than 26% improvement compared with algorithms such as SSD;on the Trash＿ICRA19 dataset,the mAP@0.5 was improved by 3.6%compared with the Faster R-CNN algorithm.