Research On Low Illumination Harbor Image Enhancement Algorithm Based On Deep Learning

In recent years,as the “Belt and Road” strategy progresses,the marine transportation industry has been developing vigorously in our country,and the cargo volume of port has increased year by year.Due to the special environment of port and wharf,when working at night,due to the serious lack of natural light,moonlight and/or lighting and other complex light environment,the image captured by the device has problems such as insufficient clarity,low contrast,high noise and blurred image details,which causes human observation difficulties,seriously affects human visual perception and further image analysis,and is not conducive to computer to perform advanced visual tasks.Therefore,technical enhancement of low illumination images is necessary both theoretically and for practical production applications.Focusing on the topic of image enhancement under low illumination,this paper adopts the network model combining Retinex theory with deep learning to train and learn the paired data sets such as LOL data set and the newly built data set based on RAISE,obtains correlation coefficients,forms different weights,and processes the images in low light.This method can better strengthen the overall brightness and contrast,but it is also accompanied by edge blur,color distortion and other problems,and the noise reduction effect needs to be further improved.To overcome the issues,this article proposes improved method.The HSV model,which is closer to the visual characteristics of human eyes than the RGB model,is used to map RGB space to HSV space,and the image is divided into three separate components,H,S and V,so as to reduce the correlation between channels and process the V component separately,which can effectively solve the problem of color distortion.Most harbor images contain ships,buildings,etc.,with complex backgrounds and high requirements for details improvement,the paper uses the attention module,to make the network focus more on improving image details and effectively reduce noise.Since the V component single channel image is processed separately,the weight influence of other channels can be directly ignored,and the spatial details can be directly paid attention to,which reduces the factors affecting the image quality.To examine the image enhancement ability of the algorithm in this paper,subjective and objective evaluation indicators are used to measure the image enhancement effect.Objective evaluation indicators: peak signal-to-noise ratio(PSNR),structural similarity(SSIM),Entropy,NIQE.According to the experimental results,the four indicators of PSNR,SSIM,Entropy and NIQE in the image processed by the improved algorithm in this paper are mostly better than other methods.Subjective evaluation indicators:overall visual effect,enhancement effect of areas of interest,authenticity of colors.By comparing the images enhanced by the improved algorithm in this paper with the images enhanced by other methods,it can be seen that the improved algorithm can effectively improve the overall brightness of the image,enhance the details of the image,reduce the noise,maintain the consistency with the original image color,and improve the visual effect.