| In recent years,the importance of underwater image processing technology in the fields of ocean engineering and underwater robots has attracted widespread attention.However,due to the absorption and scattering effects of light during underwater propagation,the quality of underwater images is severely degraded,making underwater visual detection tasks very difficult.Existing underwater image enhancement methods need to be further improved in terms of color correction,detailed texture feature enhancement,and contrast enhancement.In order to better conduct underwater scientific research,underwater image enhancement has important research value.This paper mainly studies two types of water degradation images with color distortion and unclear detail texture.According to the characteristics of the two types of images,a method of generative adversarial networks is used to enhance the distorted image.The paper is divided into the following two parts:(1)To solve the problems of color distortion and low contrast in underwater images,a feature attention network based underwater image enhancement algorithm is proposed.A feature attention module is designed in the encoder to improve the pixel feature extraction ability of the network.An adaptive feature fusion module is used in the decoder to fuse and enhance the extracted pixel features to achieve color correction and contrast enhancement of the image.In order to solve the unstable situation of network training,Spectral Normalization(SN)is introduced into the discriminator,and a dual residual module is added to the feature conversion module to reduce the amount of network parameters and make network training faster.In order to verify the enhancement effect of this algorithm,a comparative experiment was conducted using a real test set and a synthetic test set.The average value of the Underwater Image Quality Measure(UIQM)on the real test set increased by 0.155 compared to the optimal result in the comparative algorithm,and the UIQM index increased by 0.091 on the synthetic test set.The effectiveness of this algorithm in color correction and contrast enhancement is further verified through ablation experiments,RGB histogram experiments,and application testing experiments.(2)Aiming at the problems of blurred texture features and low resolution of underwater images,a super-resolution underwater image enhancement algorithm based on SK(Select Kernel)attention mechanism is proposed to enhance the detailed texture of color corrected images.The algorithm uses the global spatial attention module to extract more global features in the feature conversion stage,making the image global feature information richer.After the global spatial attention module,a SK attention mechanism is added.The SK attention mechanism can fuse different weight information based on images of different scales to achieve image detail texture enhancement,while also improving image resolution.In addition,a multi-modal loss function is designed to optimize network training,making the image perform well in terms of color and content structure.Comparative experiments were conducted using a reference test set and a non reference test set.The average value of the UIQM index on the reference test set increased by 0.107 compared to the optimal result in the comparison algorithm,and the UIQM index increased by 0.112 on the non reference test set.The ablation experiment further verified the effect of each module on image detail texture enhancement.Finally,the SIFT feature point detection algorithm and the Yolov5 algorithm were used for application testing.The accuracy and quantity of aquatic products identified by the Yolov5 algorithm were significantly improved,indicating that the algorithm can meet the needs of underwater real-time applications. |
