Research On Low Quality Image Enhancement Based On Feature Fusion Network

Image enhancement is an important research topic in the field of computer vision and image processing.Due to the different environment,the images captureed by the image acquisition equipment under the conditions of night,backlight,haze and underwater light are usually low-quality images which are damaged in some extent.The enhancement of low-quality images can improve the overall color of an image and highlight the detailed features of objects in an image,thus improving the image quality and providing effective data quality assurance for subsequent advanced visual tasks.This thesis mainly studies low-light images in natural scenes and low-quality images in underwater scenes.According to the different characteristics of these two types of low-quality images,the main achievements are as follows:For low-light images in natural scenes,a residual network based on multi-level background fusion is proposed to enhance an image by combining the image content of different levels with the corresponding background information,so that the environmental semantic information of an object can be referred to in the process of its enhancement.Moreover,a compound loss function is used to constrain the enhancement process of low-light image from three aspects of color,structure and smoothness,which makes the enhancement result closer to the real image and avoids the local over-enhancement phenomenon commonly seen in low-illuminance image enhancement.Experiments on the LOL,LIME and DICM datasets show that the low-light image enhancement algorithm proposed in this thesis can effectively improve image brightness and avoid local over-enhancement or distortion.Compared with other natural low-light image enhancement algorithms,the enhancement results of the proposed method have a performance improvement of more than 4% in terms of objective quality evaluation indexes.For degraded images in underwater scenes,a multi-input feature fusion network is proposed.Underwater images are preprocessed by three different classical algorithms to obtain three corresponding preprocessed images.The underwater images and three preprocessed images are used as the input of the network to compensate for the loss of color channels in underwater images.In addtion,the weighted idea is introduced to the color loss of the three composite losses to deal with the characteristics of the underwater image with huge color deviation due to the different distortion of different channels,so that the enhanced image has better performance in the color distribution.Experiments on the EUVP and UFO-120 datasets verify that the underwater image enhancement algorithm proposed in this thesis can effectively restore the true colors of underwater images and ensure the clarity of image details.Compared with other underwater image enhancement algorithms,the enhancement results of the proposed method have a performance improvement of more than 5% in objective quality evaluation indexes.