Underwater Low-Light Image Fast Enhancement Method Based On Deep Learning

In recent years,the ocean has become the center of human resource exploration,due to the richness and diversity of marine resources.At present,people mainly use underwater robots and machines equipped with image equipment to take underwater images and/or videos to obtain information related to marine resources.However,the imaging environment underwater is vastly different from that on land,as light undergoes scattering and refraction in water,resulting in degraded underwater images that appear blue-green and blurry In addition,since marine resources are often distributed in deep sea areas where natural light cannot reach,underwater equipment often needs to introduce artificial light sources to caputre underwater images,which results in uneven illumination distribution of images.As a result,underwater low-light images are characterized with color degradation and uneven illumination and are difficult to directly apply for marine resource detection and other advanced tasks,such as marine biological detection and recognition.In this context,enhancing underwater low-light images is essential for marine exploration and is one of the necessary steps for carrying out a series of advanced tasks.Therefore,developing effective methods for enhancing underwater low-light images is of great significance for marine resource detection.In order to effective carry out underwater tasks,there are currently two main approaches to address the degradation of underwater low-light images: one is to deal with it from the perspective of solving color deviation;the other is to solve the problem of uneven illumination under water.Current research on underwater image enhancement mainly focuses on color correction and contrast enhancement,but lacks a coordinated solution for uneven illumination and color cast problems of underwater low-light images.In order to solve these two problems simultaneously and help restore image details,this work has proposed underwater low-light image enhancement models,LUcobe and FULIE-Net.The LUcobe model is used to improve the visual display effect of underwater low-light images while the FULIE-Net built upon LUcobe achieves a balance between processing speed and image quality.Specifically,this research was carried out from the following two aspects:(1)An underwater low-light image enhancement network,called LUcobe,has been proposed to address two degradation issues in low-light underwater environments:uneven illumination and color distortion.The network comprises two components: the brightness estimation sub-network and the color correction sub-network.The brightness estimation sub-network uses an encoder-decoder structure to estimate the global brightness of the image by scaling the input image to a specific size,while the color correction sub-network adopts a dual-color space combined with an attention mechanism,which adaptively extracts the most discriminative color features,integrates them together,and finally outputs the enhanced image by combining the outputs from the brightness estimation sub-network in the channel selection module.The experimental results show that LUcobe can effectively improve the uneven illumination and color deviation of deepsea images,and achieve better visual effects and objective evaluation indicators.PSNR、SSIM 、 UIQM and MS-SIM have increased by 19.7%,12.9%,11.1% and 25.6%respectively compared with the underwater degraded images before enhancement.(2)A fast method for enhancing underwater low-light images,called FULIE-Net,is proposed based on knowledge distillation.The aim is to solve problems caused by excessive convolution such as image blurring and large model size that cannot handle real-time processing well.First,knowledge distillation is performed on the LUcobe Student network based on the FPN network structure,whose color correction part consists of the main body of LUcobe’s color correction sub-network guided by intermediate feature maps from the Teacher network.Then,an illumination loss is designed to solve the problem of uneven illumination.Experimental results show that the performance of FULIE-Net model in terms of PSNR and SSIM is close to the Teacher network model,and better in BRISQUE than the teacher model.After repeated tests of multiple images,the processing speed of FULIE-Net is nearly 40 times higher than that of the teacher model LUcobe in the model inference time performance index.