Convolutional Neural Network And Attention Mechanism-based Comprehensive Enhancement For Underwater Image

By virtue of the advantages of undertwater optical images including strong capacity in information carrying,high observability,and low acquisition cost,it plays an irreplaceable role in many fields related to ocean exploration and development.However,the underwater light propagation characteristics seriously affect the quality of images featuring in low contrast,color distortion,blurred details and other issues.By focusing on the mixed noise,light intensitydependent attenuation,domain shift characteristics in the underwater optical imaging process,this thesis solves the problem of comprehensive degradation of underwater optical images,and thereby improving visual quality and information expression ability of underwater images.The main works are as follows:Firstly,to suppress the mixed noise caused by underwater light back scattering propagation,an underwater attentional generative adversarial network is proposed.On the one hand,by virtue of dense concatation and global pooling,cascade dense-channel attention is designed to learn the discriminative features of noise,then channel-wise weight can be redistributed based on statistical probability,therey realizing the suppression of noise features;On the other hand,position attention is devised to improve the ability of convolution in capturing long-range dependent information,thereby effectively controlling the degree of image enhancement and avoiding over-enhancement.Comparative experiments show that the proposed method has outstanding noise suppression ability.Compared with full data-driven Water Net,UGAN and UWCNN schemes,the peak signal-to-noise ratio of enhanced images can be improved by about14.9%,8.4%,and 2%,respectively.Secondly,to compensate the wavelength and scene depth-dependent attenuation of underwater light,a transmission-guided dominant-features fusion network is proposed.Specifically,the channel fuzzy compensator is established to compensate the channel with the most serious attenuation,such that the influence of wavelength dependent attenuation can be efficiently relieved;At the same time,a dense multi-scale connection network is created to estimate the scene depth,such that the weight of attenuation degree about the depth dependence can be obtained;Consequently,different distribution features are integrated,and the dominant features of three branches’ input are fused under the guidance of attenuation degree weight to achieve image enhancement.The comparisons of feature expression ability show that the proposed method can significantly improve the ability of image feature expression.Compared with original image,the results of enhanced image in SIFT feature point matching,Harris corner and Canny edge detection are increased by 16.8,7.8,and 25.8 times,respectively.Finally,to suppress the domain offset caused by the inconsistent degradation characteristics of underwater scene,and break the imbalance between image visual quality and semantic information expression ability,an underwater object detection-driven image enhancement network is proposed.On the one hand,a codec image enhancement network based on residual structure is created to realize the reflection from degraded images to enhanced images;On the other hand,the object perceptor is pre-trained to extracte the semantic information and predict location boxes and labels;Combined with joint loss and multi-stage loss strategy,the enhancement network can be driven to generate images with both visual quality and semantic information expression ability.From the comprehensive comparisons of image enhancement performance and semantic information expression ability experiments,the proposed image enhancement scheme has the best balance ability in terms of visual quality and object detection accuracy.And the preprocessing effectiveness experiment shows that the designed network can effectively suppress the decline of object detection accuracy caused by domain offset,and thereby increasing the AP values of sea cucumber,scallop and starfish categories by 15.6%,4.5%,and 11.1%,respectively.