Research On Infrared Guided Low-light Image Enhancement Technology Based On Deep Learning

Low-quality images can include some valuable content,which may be difficult to understand due to the poor quality brought by the limited capture environment,and thus also inhibit their applications to image processing and computer vision.This thesis focuses on the image enhancement of such low-light images.Some existing cameras,especially monitoring ones,have the infrared mode with the near-infrared image not affected by the low-light environment.Therefore,we take infrared information and construct deep neural network models to repair the low-light image.The main contents of this article include:First,the related researches are investigated first,including traditional and deep learning based low-light image enhancement,and then the background knowledge and related theories are introduced.Secondly,a full convolution depth low-light image enhancement model with infrared information is proposed to solve the missing problems in the low-light images.It uses the endto-end modeling of the codec network based on multi-scale features to model the mapping between low-light images and normal brightness images,and merges the scene structure in infrared images to guide the enhancement of low-light images.The pixel-level content difference between the generated image and the target image during training,the structural depth feature difference and the loss of auxiliary infrared information at the decoding end are taken as constraints,so that the performance of low-light image enhancement is improved,which demonstrates the effectiveness of incorporating the infrared information.Finally,a deep anti-low-light image enhancement model based on the attention of infrared features is proposed to tackle that local area color distortion existing in the output of the proposed full convolutional network.It includes two sub-networks,a generative model and a discriminant model: the former uses a context coding method to build a situational focus network based on the correlation of regional features,optimizing the local area and generating a new repair image;the latter determines the authenticity of the generated image by checking the result of low light image enhancement.The scene attention layer evaluates the structural correlation between local regions based on the characteristics of the infrared image and thus can overcome the instability of the generated image during the optimization process.Experiments verify the effectiveness of the two-stage model.