Research About Low-illumination Image Enhancement And Super-resolution Reconstruction Algorithm Based On Convolutional Neural Network

Due to the weather,illumination and the quality of image acquisition equipment,it is often difficult to obtain the expected image quality.The cost will be too high by replacing the imaging equipment,and it will still be affected by the environment and climate.While improving the image quality through software algorithm can greatly reduce the cost.In order to improve the image quality,from the perspective of image brightness and resolution,this paper proposes a low illumination image enhancement algorithm based on improved attention mechanism to enhance the brightness of the image,and an image superresolution reconstruction algorithm based on multi-scale fusion convolution neural network to improve the image resolution.The main research work and results are as follows:(1)Aiming at the problems of low brightness,poor contrast,fuzzy image information and difficult to distinguish the image target in the image,a low illumination image enhancement algorithm based on Retinex theory and attention mechanism is proposed in this paper.Firstly,an improved attention mechanism is added to the decomposition network to pay attention to the key information in the image and allocate computing resources reasonably,so as to achieve better decomposition effect;Secondly,by using reconstruction loss,constant reflectivity loss and illumination smoothing loss,the decomposition and enhancement ability of the network are optimized to improve the quality of the generated image.Experimental results show that compared with other algorithms,the proposed algorithm achieves the best results in multi angle comparison.(2)Aiming at the problems of insufficient feature extraction and artificial redundant information caused by deconvolution in FSRCNN model,an image super-resolution reconstruction algorithm based on multi-scale fusion convolutional neural network is proposed in this paper.Firstly,a feature extraction channel of multi-scale fusion is designed to solve the problem of insufficient utilization of different image size information;Secondly,in the part of image reconstruction,sub-pixel convolution is used to suppress the checkerboard effect and achieve better reconstruction effect.Compared with FSRCNN model,in set5 and set14 data sets,the PSNR and SSIM values under 2 times amplification factor are increased by 0.14 db and 0.0039 on average,and 0.48 db and 0.0072 under 3 times amplification factor on average.Through the analysis of the experimental results,this algorithm can not only retain more image details,but also improve the overall image reconstruction effect.(3)Based on the two algorithms proposed in this paper,an interactive image processing system is designed and implemented.The system can not only improve the brightness of dark images and low illumination images,but also improve the resolution of blurred images and get clearer images,so as to improve the total body mass of images.By improving the quality of various images,the system meets the needs of the society to obtain high-quality images,and realizes the application value of the algorithm proposed in this paper.