Research On Low Illumination Image Enhancement Method Based On Retinex And Convolutional Neural Network

As an important carrier for people to communicate with each other,visual images have become one of the important ways for people to acquire and exchange information.With the continuous development of computer vision,machine learning and artificial intelligence technology,images are widely used in video surveillance,medical diagnosis and treatment,radar remote sensing detection and recognition and other fields.Yet,due to the limitations of image acquisition equipment and the influence of external ambient light,the images collected by people often have problems such as low contrast,serious color deviation,loss of detail information and so on,which seriously hinders the subsequent application of images.Therefore,the research on how to achieve low illumination image enhancement and obtain high-quality image has important research value.Although the existing low illumination image enhancement methods have achieved certain research results,the loss of image details and color distortion will still occur in the face of local dark or information rich low illumination images.Based on the above problems,this paper studies the low illumination image enhancement methods.The main work and contributions of this paper are as follows:(1)In this paper,the relevant theoretical knowledge of low illumination image enhancement methods is discussed in detail,and two types of classical low illumination image enhancement methods will be deeply studied,code reproduction and experimental verification of related methods will be carried out,and the advantages and disadvantages of these methods will be analyzed,and the key problems that can be used for reference and need to be paid attention to in the process of low illumination image enhancement will be summarized.(2)Aiming at the problem that the overall brightness of low illumination image is too low or local brightness is too dark,resulting in the loss of detail feature information and blurred image edge detail,a U-net-based multi-scale feature preserving method for low light image enhancement is proposed in this paper.Inspired by Retinex basic model theory and U-Net network structure,combined with multi-scale extraction technology,this method further improves the feature extraction ability of the network.The discontinuity between low-level features and high-level semantic features is solved by maintaining the spatial consistency function of global feature correlation.At the same time,a multi-scale structure calculation method based on image brightness,contrast and structural similarity is introduced,which greatly alleviates the phenomenon of uneven illumination and color deviation after enhancement,and makes the enhancement result more in line with human visual perception.Extensive experiments show that compared with other advanced enhancement methods,this method has better enhancement effect and can retain more details.(3)Aiming at the problem of insufficient detail enhancement and color distortion of low illumination image in extreme environments,a low illumination image enhancement method based on U-shaped multi-scale network is proposed in this paper.This method improves the feature extraction ability of the input image by the decomposition network by combining the parallel multi-scale feature extraction module with the three-layer U-shaped network structure.The multi-scale network block is introduced to replace the single convolution layer,and the U-shaped network structure is combined to realize the fusion of context information,which further reduces the problem of information loss in the process of image enhancement.At the same time,the color loss and Multi-Scale Structural Loss functions are introduced to constrain the color information of the image,so that the output image is closer to the image in the nature environment.Experimental analysis results also show that the method proposed in this paper has excellent performance in low illumination image detail information extraction and color correction,and has a good enhancement effect.