Research On Low Illumination Image Enhancement Algorithm Based On Deep Learning

In recent years,with the wide application of computer vision technology,image enhancement has become more and more important as a basic research work in the field of computer vision,among which low illumination image enhancement is one of the important research topics in the field of image enhancement.At present,various lowillumination image enhancement methods based on different theories,as well as various deblurring,denoising and brightness adjustment technologies have been proposed,but the image enhancement technology still faces many challenges.In addition,low illumination images are often accompanied by image degradation and lack of detail,which also brings a lot of difficulties for image restoration.Aiming at the problems of image noise removal,color distortion and image detail recovery in low illumination environment,the research of image enhancement in this environment has certain application valueIn this thesis,the image enhancement problem in low illumination environment is studied,and the main work is as follows:1)The data set of outdoor multi-exposure low-illumination image was constructed.In this thesis,the existing low-illumination data sets are analyzed,and the outdoor lowillumination data sets are constructed by a variety of shooting methods using multiple imaging devices such as cameras and smart phones to increase the diversity of the data sets in real scenes and ensure the robustness of subsequent algorithms.2)in combination with Retinex theory,this thesis proposes an end-to-end improved convolutional self-coding network structure,which consists of an encoder and a decoder.Considering the challenges of noise removal,color distortion and local detail restoration,three network modules of image decomposition,illumination adjustment and image reconstruction are designed.The network encodes the image through the image decomposition module and obtains the reflection feature encoding and illumination hiding encoding.Among them,the network illuminance adjustment module automatically adjusts the image illuminance,the image reconstruction module realizes the restoration and reconstruction of the image content,and obtains the final enhancement result.3)because the image decomposition based on Retinex theory is an ill-posed problem and has no reasonable constraint on the output,this thesis designs a joint loss function to constrain the training of the network.The loss function is divided into three parts: reconstruction loss,color loss and reflection loss.In this thesis,the reflection loss constraints,including detail constraints and structural similarity constraints,are designed based on the prior knowledge that the reflection features of low-illumination images and normal exposure images should have the same structure and details.At the same time,in order to ensure that the content of the reconstructed image is closer to the image with normal exposure and to correct the color deviation,the reconstruction loss and color loss are designed.Finally,the effectiveness and advancement of the proposed algorithm are verified by the experimental results.