Research On Underwater Image Enhancement Algorithm With Dual Branch Cascade Based On Deep Learning

Underwater image enhancement techniques have important research implications in the field of underwater robotics and marine engineering.In recent years,underwater image enhancement and recovery algorithms have attracted an increasing amount of research work and play a crucial role within the field of computer vision.In underwater scenes,the visibility of images and videos is often reduced due to wavelength-dependent light absorption and scattering by suspended particles in the water.Examples include blue-green underwater images due to light attenuation;blurred underwater image sharpening due to forward scattering of light and low contrast problems due to backward scattering.Degraded underwater images and videos affect the accuracy of underwater scene pattern recognition,visual understanding and key feature extraction.It poses a challenge for marine engineering and underwater exploration tasks.In order to solve the above problems,this paper designs and investigates a dual-branch undewater image enhancement algorithm based on media transfer map guided by an efficient recovery Transformer(Restormer)based multi-scale and multi-patch feature fusion cascade guided dual-branch underwater image enhancement algorithm and.The specific studies are as follows:(1)Underwater image enhancement algorithm guided by a medium transport map is improved.To address the problems of low contrast and inaccurate colour correction in the enhanced results of this algorithm,this paper firstly extends the medium transfer map module by increasing the number of convolutional layers to enhance the response of the network to degraded regions and extract more image features.Secondly,a multi-headed attention mechanism is introduced to focus on the information between different channels to solve the problem of noise present in the images.Finally,the dilation rate of the dilated convolution is changed so that it is arranged in ascending and descending powers in equal proportion to reduce the redundancy present in the network space.The results show that the improved method has significantly improved both the subjective visual effect and the objective quality evaluation.(2)To address the problems of inaccurate feature extraction and blurred details of underwater images of different scenes by current underwater image enhancement algorithms.In this paper,we propose the Res-DNet underwater image enhancement algorithm,a two-branch cascaded feature fusion network based on an efficient Restormer model.The algorithm combines multi-scale(MSHN)and multi-patch network(MPHN),aiming at correcting colour bias,improving contrast and deblurring details in non-uniform underwater images,and recovering non-uniform underwater images by aggregating features from multiple scales in different spatial regions with a small number of network parameters.We design a three-level cascade architecture to enhance the generalization capability of the network,aggregating the learned features from two branches by jumping connections,and adding another Restormer at the end of the network as a refinement network to enhance the network performance by a linear weighted combination of each loss.The effectiveness of the algorithm in this paper is verified through experiments.(3)As with the other methods,we tested on the Test-R90,challenge-60,Color-checker7,and SQUID datasets,comparing both subjectively and objectively with the other eight advanced methods.The experimental results show that our method is more applicable to different underwater scenes and shows significant improvements in colour correction,contrast enhancement,edge The results show that our method is more applicable to different underwater scenes and provides significant improvements in colour correction,contrast enhancement and enhanced edge information.In addition,we synthesise underwater image datasets to enhance our network training.Finally,we validate the effectiveness of the modules in the network through ablation experiments.(4)For the method proposed in this paper,we applied it to saliency target detection in underwater scenes,performing target detection on real-world underwater images and enhanced underwater images respectively,and found through experiments that our algorithm is also applicable to other vision tasks.