Research On Low-Light Image Enhancement Method Based On Deep Learning

When using a photosensitive device to capture images,it is usually encountered that the lighting conditions are insufficient or the exposure time is short,which will cause the captured images to have a lower grayscale resolution.This type of image is generally called a low-light image,which not only brings a very poor perception to users,but also interferes with further applications in criminal investigation,medicine and other fields.At this time,if the low-light image enhancement algorithm is used to process the low-light image,the image quality can be improved from the aspects of contrast,signal-to-noise ratio,color correction,etc.,so as to meet the subsequent application requirements of the image.Based on the analysis of existing low-light image enhancement algorithms,this thesis proposes two lowlight image enhancement algorithms based on deep learning,one of which is an end-to-end enhancement network,and the other is an enhancement network based on Retinex theory.Both have achieved relatively good subjective and objective enhancement effects.The main research work of this thesis is as follows:(1)A low-light image enhancement algorithm based on selective fusion of multi-scale features is proposed.Aiming at the disadvantage that traditional algorithms cannot extract image features from multiple angles,this algorithm uses the Attention U-net structure with a gated attention mechanism to extract features from images,allowing the network to selectively extract image brightness from multiple angles Features and details.Aiming at the problem of reduced generalization ability of convolutional neural networks due to excessively deep layers,the algorithm draws on the multi-scale fusion idea of the MSR algorithm,and extracts middle and low-level features from the above Attention U-net structure as multi-scale features for fusion.The network fully explores the connections between the features of adjacent layers,thereby improving the generalization performance of the network.In order to overcome the disadvantage that the MSR algorithm needs to artificially set the multi-scale fusion weights and make the algorithm not adaptive,this algorithm uses the SE channel attention module to adaptively learn the multi-scale fusion weights,so that these multi-scale features can selectively interact with each other.The output features of the Attention U-net structure are added and fused.Experimental results show that compared with other algorithms,this algorithm achieves better results in overall image contrast stretching and detail preservation.However,the algorithm in this chapter still has the blindness of end-to-end network extraction of image features,which will cause image blur and color distortion in the enhancement result.(2)A low-light image enhancement algorithm based on Retinex theory and illumination guidance is proposed.The algorithm first extracts the reflection component and illumination component of the dark light image through the illumination decomposition sub-network,and then outputs the final enhancement result through the enhancement result prediction subnetwork.Aiming at the shortcoming that the existing illumination decomposition network based on Retinex theory is not stable,the illumination decomposition network of this algorithm combines the prior knowledge of illumination to only extract the illumination component of the input image,thereby reducing the difficulty of network training.The existing algorithms based on Retinex theory adopt separate optimization strategies for the extracted illumination components and reflection components,which will cause loss of image information.Aiming at this problem,the enhancement result prediction sub-network of this algorithm uses a two-stage loss function to optimize the illumination component and the reflection component at the same time,so as to avoid the separate optimization strategy to amplify the noise.Aiming at the problem of a large amount of noise and color distortion in the reflection component,the algorithm proposes a Dense module based on dilation residuals,which enables the enhanced result prediction network to extract noise and color features from different receptive fields of the reflection component.In order to alleviate the problem of image blur caused by uneven illumination,the algorithm uses an illumination fusion module based on an improved gated attention mechanism,which enables the enhancement result prediction network to adaptively enhance image details under the guidance of illumination.The experimental results show that,compared with the end-to-end network proposed in(1),this algorithm effectively solves the problem of image blur and color shift in the enhancement result by combining the prior knowledge of the image,and obtains better results with less computational complexity.It is an excellent enhancement effect.