Deep Learning-based Underwater Image Enhancement And Object Detection

Underwater image enhancement and target detection are the key technologies for human exploration of the ocean,and are widely used in underwater fishing,underwater survey and other fields at home and abroad.At present,image enhancement and target detection technology in complex underwater environment have always been research hotspots.This thesis comes from the National Key R&D Program "Ship-borne Unmanned Submersible Retraction System"(2018YFC0309100),and focuses on image enhancement and target detection technology in harsh underwater environment.And the main research work is as follows:Acquisition of underwater images and data set expansion.An underwater experiment platform is built,and underwater targets are selected.Then target image samples are collected by using the underwater experiment platform.Finally,a method based on the fusion of Mixup pixel features is proposed to expand the collected underwater image data set for subsequent underwater image enhancement and underwater target detection.Research on underwater image synthesis algorithm.Aiming at the problem that the underwater scene of the current underwater-land paired data set is not comprehensive,an underwater image synthesis algorithm that integrates underwater imaging model and generative adversarial network is proposed.The terrestrial image data set is used to generate underwater images in different styles to form an underwater-land paired data set that is for enriching underwater image enhancement experiments.Research on underwater image enhancement algorithm based on CU-Retinex Net.Based on the analysis of currently used underwater image enhancement methods,CU-Retinex Net is proposed to solve the problems of poor contrast,unclear details and image color cast of underwater images based on Retinex Net.The proposed algorithm decomposes the original image into illumination and reflection component images by combining data-driven convolutional neural network(CNN),adopts a U-Net network that fuses residual blocks to improve the brightness and contrast of the image from adaptation,retains the detailed information of the image,and finally multiplies the denoising reflected component image with the illumination component image element by element,and finally obtains a high-quality underwater enhancement image.The experimental results indicate that CU-Retinex Net can effectively improve the brightness,contrast,edge details of the image,and restore the color realism of the image.Research on underwater target detection algorithm based on DDA-RetinaNet.Based on the analysis of commonly used deep learning target detection methods,DDA-RetinaNet is developed to balance the detection speed and detection accuracy in underwater target detection.The proposed algorithm integrates the Dense Net structure that can learn group convolution to improve the extraction ability of features.Dual-FPN is used to improve the detection capability of the network on overlapping underwater targets.The Anchor Free-based method is used to improve the detection ability of underwater small targets while reducing the adjustment of model parameters.Through the subjective and objective data of the experimental results,it is proved that the algorithm can effectively enhance the robustness of underwater target detection,and can meet the requirements of real-time detection while ensuring the detection accuracy.