Research On Low Illumination Image Enhancement Based On Retinex Theory

In the process of image acquisition,there are often some uncontrollable factors,especially in the case of poor lighting conditions such as indoor,night or cloudy day,which will lead to various defects in the collected image.The image quality of this low lighting image may be seriously degraded due to color distortion and noise,which will not only affect human visual senses,but also affect other types of computer vision tasks.In conclusion,it is valuable to study the direction of low illumination enhancement in the field of image enhancement.Focusing on low light image enhancement,this paper studies and analyzes the current mainstream models,and finds that there are problems such as multi noise and color degradation after image enhancement for a single low light scene,while there are problems such as over exposure and over saturation after image enhancement for complex low light scenes.The problem of low illumination is solved in this paper.What has mainly investigated in this paper are present as follows:1.Aiming at the problems faced by low illumination image enhancement,such as difficulty in selecting reference standard image and image detail degradation.In this paper,an improved kind low illumination image enhancement model is proposed.Based on Retinex theory,the network structure is divided into three parts: decomposition network,illumination enhancement network and reflection enhancement network.Hoping to solve the problems of detail enhancement,color restoration and color block removal in low illumination images,a residual dense denoising module is proposed in the decomposition network to effectively solve the problem of multi noise in the decomposed reflection component images;In the illumination enhancement network,a new codec network structure is proposed,which combines the local smoothing loss function to obtain the lower level feature information and effective constraint enhancement results;In the reflection enhancement network,combined with the hybrid domain attention mechanism,further texture restoration is done for the local region.2.Based on the theory of low illumination and non-uniform illumination,this paper still puts forward the problem of image enhancement based on low illumination.In this model,in the decomposition network,the illumination component image is estimated to produce the channel initialization mapping result,which is convenient for the subsequent enhancement tasks to a certain extent;A new loss function is designed in the illumination enhancement network for the uneven illumination distribution in uniform low illumination images;A new fusion module is designed in the reflection enhancement network to effectively fuse the extracted multi-scale feature information.The two models proposed in this paper have carried out a large number of comparative experiments and ablation experiments on multiple mainstream data sets.Comparative experiments show that the recommendation performance of the two models proposed in this paper has a certain improvement in the benchmark model.In addition,ablation experiments also verify the effectiveness and rationality of the design of each module in the model proposed in this paper.