Research On Low Illumination Video Image Enhancement Technology

With the development of society and hardware equipment,some have collected more and more video images,and various images have enriched our lives,but the quality of a heavy number of images is poor,especially in underexposed environments such as dusk and night.The acquired images are often of low quality,so that the relevant researchers cannot extract the necessary information in the images,which needs to be processed and used as the basis for the next one.Therefore,it is of nice practical significance to enhance the low-illuminance video images.The contribution of this subject is summarized as follows: First of all,we elaborated on the research background and significance of the subject,while analyzed the domestic and foreign research status,development tendency and problems of low-illumination video image enhancement at hand.Research and simulation of low-illumination video image methods of some scholars are performed by us,such as histogram-based method,the methods of Retinex and deep learning,were conducted to understand and analyze their advantages and drawbacks.Then in view of its shortcomings,two methods are proposed to enhance low-illumination images based on the previous analysis of other low-illumination video image enhancement methods:(1)We propose an improved algorithm based on the Retinex algorithm,as for the halo artifacts and loss of details after the traditional algorithm enhances the low-illuminance image.First,the low-illuminance image is converted from the RGB image to the HSI color space,Only the luminance component I is processed,the guided component filtering of the fusion edge detection operator is used to convolve the I component to obtain the illuminance image,then the illuminance image is removed to obtain the reflected image,next the low rank decomposition and local contrast enhancement are introduced to further improve the image,the last one converts the HSI image to RGB image in order to have the final enhanced image.The algorithm in ours can keep balance in restoring details,suppressing noise and removing artifacts.(2)We propose a low-illumination image enhancement algorithm based on convolutional neural networks based on deep learning methods,aiming for the fact that traditional algorithms require human intervention in the enhancement process and it is hard to find the constraints of various scenarios.First,the low-illuminance image is decomposed in a data-driven manner,into an illuminance image and a reflection image.The illuminance image adopts the Encoder-Decoder network structure,and the illuminance image is adjusted with the decomposed reflection image;the reflection image is also enhanced by the convolutional neural network;eventually,the two enhanced images are combined into an enhanced image according to the Retinex pipeline.The algorithm can enhance the image of a large area of dark area and obtain a higher quality image effectively.Finally,a system design was carried out in order to verify the practicability of the algorithm in this paper.On the basis of method 1,a low-illumination video image enhancement system was designed and implemented.The system mainly processes images and videos,and users can target them according to their own needs.Low hardware requirements,simple interface,and is easy for users to operate.Experiments on the system show that the system can better enhance the video image and meet the needs of users basically.To sum up,two algorithms are proposed for the deficiency of low-illumination image enhancement methods,which are improved on the basis of Retinex and deep learning,and well experimental results are obtained.Finally,the improved Retinex algorithm is designed as a low-illumination video image enhancement system,and the single image and video are enhanced simultaneously,which can improve the image visual effect.