Research On The Low-light Image Enhancement Based On Deep Learning

In low-light environments,the images captured by imaging equipment often have many problems such as loss of detailed information,low contrast,difficult protection of dynamic range,and a lot of noise.Such images will not only affect human's ability to visually perceive images,but also affect the performance of artificial intelligence algorithms based on visual information.Through image enhancement technology,the quality of low-illumination images can be effectively improved,mainly in terms of better sharpness,texture detail information,and lower image noise.Low-light image enhancement currently faces many challenges and also has very important applications in human life.Therefore,this problem is huge in both research and practical applications.With combining a deep neural network algorithm that has made many advances in the field of computer vision and the idea of multigranularity analysis,this paper focuses on how to enhance the image quality in low-light environments and conducts a series of studies.The main work of the article includes:First,we introduce relevant background knowledge in chapter 1 and2.In the third chapter,based on the influence of convolution kernels of different size receptive fields in deep neural networks on the performance of deep neural networks,combined with the ideas of residual networks and dense networks,inspired by multi-granularity analysis,a multi-grained residual dense convolution block is proposed.This method uses different convolution kernels of different receptive field sizes in a module to extract information of different granularities on the features output from a layer on the deep convolutional neural network.At the end of the module,new features are generated by fusing the feature information under different granularities.Based on the multi-granularity residual dense convolution block,this paper proposes a multi-granularity residual dense network.The validity of this method is proved by comparing with the existing methods and studying the degradation of the model.Then,in the fourth chapter,based on the analysis of multi granularity theory,a multi granularity cooperative deep neural network model is proposed.This model uses multiple simple single granularity neural networks to carry out two-way information interaction through cooperative learning,so as to achieve progressive performance enhancement.Through a large number of experimental analysis and degradation research,this method has made a significant improvement in the current open data set.Finally,in the fifth chapter,in view of the problem of image dynamic range protection in the low-light image enhancement algorithm,through the study and analysis of deep neural network learning methods and mapping relationships,a light map estimation function is proposed.The input data is used for local lighting adjustment to protect the dynamic range of the image during the enhancement of the low-light image.In order to verify the algorithm,this paper collects high dynamic data sets in low illumination scenes,and verifies the effectiveness of the algorithm through experimental comparison.