Research On Enhancement Algorithms Of Low Illumination Images Based On Convolutional Neural Networks

In the case of low illumination environments such as night,the acquired images will have some problems such as narrow dynamic range,lack of detailed information,and a large number of global noises.These types of images will not only affect human visual perception ability,but also bring difficulties to the further image processing.It is obvious that the research on low illumination image enhancement has rich application value.This thesis studies many algorithms for image enhancement based on convolutional neural networks,and analyzes the images collected in indoor and outdoor scenes with uniform lighting and non-uniform lighting,and improves some of the current poor algorithms.The research contents of this thesis are as follows:1.This thesis proposes an enhancement algorithm for uniform illumination images based on an improved Retinex-Net and illumination smoothing.It uses the Retinex-Net network as the theoretical basis for uniform illumination image enhancement,aiming at low illumination images with uniform illumination(that is,such images are mostly low illumination images of indoor scenes,the overall image is dark,and the naked eye can barely perceive the details)for research.According to the Retinex theory,this thesis first uses the illumination smoothing loss function to decompose the image into a relatively accurate illumination map and reflectance map,and finally synthesize the enhanced illumination map and the denoised reflectance map to obtain a new image with better visual perception.Experimental results show that the proposed algorithm not only effectively improves the brightness and contrast of the detailed area of the uniformly illuminated low illumination image,but also improves a lot under various objective index values.2.This thesis proposes an enhancement algorithm for non-uniform illumination images based on U-Net and attention mechanism,and uses the U-Net network and attention mechanism as the theoretical basis for non-uniform illumination image enhancement,aiming at low illumination images with non-uniform illumination(that is,such images are mostly low illumination images of outdoor scenes,which are affected by the illumination from different directions so that the brightness of each area is different,and some details can be perceived by the naked eye)for research.The algorithm uses channel attention to predict the light distribution in low illumination images with non-uniform illumination.It is expected to learn a mapping function to arbitrarily convert images with multiple light levels into images with another light level to guide enhancement the network adaptively enhances the different brightness areas in the image.At the same time,a deep U-Net network is designed to remove the noise in the reflectance map,and the attention strategy is used to suppress the undesired chromatic aberration and noise.Experimental results show that the proposed algorithm not only effectively improves the contrast and brightness of low illumination images with non-uniform illumination,but also has good effects in terms of color saturation and image noise suppression.