Research On Image Enhancement Algorithms Based On Deep Learning

Under the limitation of the natural environment with low light or the low accuracy of digital imaging equipment,the captured images usually have low brightness,high noise,and color distortion.Their resolution and dynamic range are also limited to some extent,causing visual effects problems,and can not effectively extract image information in computer vision tasks.Aiming at this problem,this paper has done the following work on image enhancement research:1.The image enhancement model based on the enhanced anti-learning Exposure framework can complete the image modification in the data set that does not correspond to the label and input.But the enhancement result has the problem of overexposure,color distortion and so on.According to the problem,we propose relativistic adversarial advantage actor-critic with critic regularization framework.In this paper,the relative mean generative adversarial network is used to approximate modeling the reward function in the reinforcement learning framework to strengthen the discriminator's discrimination ability and design the learning behavior of the penalty constraint strategy network of the strategy gradient algorithm to stabilize the training.Our method improves the quality of the enhanced image.2.Most single image enhancement methods cannot display image details due to limited original image information.We use deep fast guided filtering to decompose the image into low-frequency base layer and high-frequency detail layer,and then learn the corresponding features through residual convolutional neural network,so that the network focuses on image color conversion and restoration of high-frequency details,respectively.Finally,high-frequency and low-frequency features are combined to recover high quality images,and pixel-by-pixel enhanced images are achieved.3.For the traditional HDR image synthesis technology,it is necessary to synthesize multiple images with different exposures,and then generate HDR images through tone mapping,which is more complicated and time-consuming.So we propose an algorithm base on generative adversarial networks to transfer LDR image to HDR style.Generative adversarial networks are widely used in image editing tasks.In this paper,we use an efficient one-way generation adversarial network.The multi-scale adaptive conditional normalization module is uesd to learn the HDR style image characteristics of the target domain and generate the style affine parameters of the normalized layer to guide the implementation of global and local style transfer when feature layer reconstruct.In the decoding process,in order to preserve the structural information and statistical characteristics of the feature space,a wavelet pooling operation is used instead of the traditional pooling operation to generate a high quality HDR image.