Research On Underwater Degraded Image Enhancement And Target Detection Method

Underwater target detection based on vision has played an important role in Marine economic fields such as Marine environmental protection,water ecological research and underwater garbage cleaning.However,in the complex water environment,low contrast,blur,color distortion and other degradation problems are caused by the absorption and scattering characteristics of light.At the same time,there are many problems such as occlusion and overlap among underwater targets with different scales,which bring great challenges to underwater target detection.Degraded image enhancement can improve image clarity,subjective effect and reliability of subsequent detection tasks.Therefore,the research on underwater degraded image enhancement and target detection has a certain application prospect and research value.Therefore,this thesis proposes an underwater image enhancement method based on multipath extended dense network and an underwater target detection algorithm based on multi-scale strategy.The main work is as follows:(1)Most of the existing physical model-based CNN underwater image enhancement methods optimize the image indirectly by estimating the transmittance map and background light value.The degree of image enhancement is limited due to the influence of prior conditions and model cumulative error.To solve the above problems,this thesis proposes an end-to-end underwater image enhancement method based on multipath expansion dense network,which avoids the errors caused by the above indirect optimization and has better visual effects.Firstly,the multi-path dilation feature extraction module is used to obtain multiscale features of images,and the multi-path dilation convolution block and cascade attention mechanism are introduced to expand the receptive field and improve the overall ability of the network to perceive details and color information.Secondly,the optimize densely connected block is designed and optimized to strengthen feature transfer and aggregate feature output of joint univariate feature map to enhance the network’s ability to utilize different degraded pixel features.Finally,a clear underwater image is obtained through the underwater imaging physical model.Experimental results show that the enhanced image obtained by the proposed method has more details and higher color reduction degree.In the Trash＿ICRA19 dataset,the underwater image evaluation index UCIQE and UISM scores reach 0.597 and 4.663,and the comprehensive performance is better than other methods.(2)Aiming at the problems of different scales of underwater targets and overlapping occlusion among them,an underwater target detection model based on multi-scale strategy is proposed.The model architecture is two-stage target detection.Based on Resnet101 residual network,multi-layer feature fusion module and multi-expansion dilated convolution pyramid module are designed for feature extraction to capture more multi-scale feature information of context.Secondly,Ro I Align pooling is adopted to cancel the quantization operation,so as to retain the pixel value mapped to the feature map of the recommendation region and reduce the error.Finally,a new loss function AD-Io U is proposed to realize adaptive optimization of prediction box.The effectiveness of the proposed method was verified on Labeled Fishes in the Wild and Trash＿ICRA19 dataset.The results show that the proposed method has good scale adaptive ability under different occlusion rates,and the detection accuracy of Trash＿ICRA19 dataset is 80.1%,which is better than other models.