Research On Low Light Image Enhancement Method Based On Deep Learning

In low light environment such as night and backlight,due to insufficient exposure,there are serious noise,low information content and poor contrast in the image taken by imaging devices.These problems will not only greatly affect people’s pursuit of visual experience and daily sharing needs,but also affect the implementation of computer vision tasks such as object detection,object classification,etc.The low-light image enhancement technology can reduce the noise in the under-exposed image,improve its light intensity and make it have a high contrast,which has high research and application value.In recent years,the technology of low-light image enhancement has made great progress.There are some traditional methods such as histogram equalization,gamma correction and Retinex theoretical model,etc.Moreover,many deep learning methods based on paired data sets,non-paired data sets and single image data sets have been proposed.However,the images enhanced by existing algorithms still have low contrast,serious noise pollution,the lack of detail.Most importantly,there are few methods to consider the characteristics of the low-light image enhancement task that different individuals have different enhancement feelings.In response to the above problems,this thesis mainly carries out the following work:(1)Low light image enhancement method based on full convolution network.In this thesis,the composition and characteristics of full convolution networks are analyzed in detail.Based on Res-Net and U-Net,a new full convolution neural network is constructed,and the experimental comparison and analysis are made with the current mainstream full convolution neural networks for the low light image enhancement task.Finally,the parameters,loss function and network model that are most suitable for this task are obtained.(2)Low light image enhancement method based on Retinex theory and deep learning.In view of the problems of blurred object contour,dim color,missing details,and discontinuous color in most of the images enhanced by low-light image methods in extreme cases,this thesis constructs an image decomposition module,a reflection component recovery module and a brightness enhancement module based on Retinex theoretical model to decompose the low-light image into its reflection component and illumination component respectively.Noise removal and color correction are carried out for the reflection component,and the illumination component is enhanced.The enhanced image is obtained from the restored and enhanced reflection component and illumination component.Through extensive analysis and comparison with related representative algorithms in the same field,the effectiveness of this method is verified.(3)Low light image enhancement algorithm based on dynamic network.Aiming at the problem that traditional deep learning methods cannot meet different subjective needs of different users in low-light image enhancement tasks,this thesis designs a dynamic training strategy that uses different data sets to train different parameters in stages,and constructs a deep learning network that users can adjust the enhancement effect according to their subjective preferences.Subjective,objective,and generalized experiments were made on images from different data sets and with different brightness levels,and volunteers were invited to evaluate the enhancement results of various methods and select the image that best suits their aesthetic preferences.The experimental results demonstrate that the dynamic network in this thesis can well meet the user’s subjective preferences.