Image Enhancement Based On Multi-Scale Convolutional Neural Network

Image enhancement algorithms can obtain higher-quality images without relying on hardware,and can also solve the problem of image transmission.In recent years,image enhancement has become a research hotspot with many new and practical algorithms being appeared.Among them,image enhancement based on deep learning(especially convolutional neural networks)can significantly improve image quality,which has attracted widespread attention from researchers.However,the current image enhancement algorithms based on deep learning are mostly single-scale models,which cannot effectively capture robust features to deal with scale transformation problems.At the same time,the model with large volume of parameters is difficult to deploy effectively for applications.This thesis exploits issues in single-image super-resolution reconstruction and remote sensing image enhancement.The research of image enhancement algorithms based on multi-scale convolutional neural network is carried out,mainly including the following two aspects:First,based on the summary of past classic deep learning super-resolution reconstruction models,this paper proposes a novel multi-resolution convolutional neural network super-resolution reconstruction model based on hierarchical fusion and splitting ideas,namely RFNet(Residual Fusion Network).Considering the scale transformation problem in the image enhancement(super-resolution),a multi-scale feature extraction module is introduced,which makes the model robust to different scales.At the same time,a framework with hierarchical split and fusion is introduced into the network,which further compresses the model parameters.The super-resolution reconstruction experiments were carried out on the computer vision simulation data set and the real data set separately.The results verified the validity of the RFNet model accordingly.Compared with the existing deep learning models(all methods set the same number of layers),the RFNet model has advantages in various image quality evaluation indicators,and the parameters of the model are more streamlined.Second,the applicability of RFNet in optical remote sensing image enhancement(Pansharpening)was discussed.A modified version of RFNet(M-RFNet)was proposed mainly according to the characteristics of the remote sensing data.The M-RFNet model is highlighted mainly in three aspects:(1)in terms of the model complexity,the size of the entire model is compressed to prevent overfitting problems;(2)the perception field of the network is further increased to obtain more feature information;(3)the combination of convolution with PixelShuffle is used rather than the bicubic interpolation,and the up-sampling operation through network autonomous learning is implemented.Experiments based on Gaojing(SuperView)-1 image data showed that an improved image enhancement was obtained through M-RFNet model.Compared with the traditional Pansharpening methods,the M-RFNet model can achieve better image fusion results with less spectral and spatial distortion.In addition,the experimental results demonstrate the importance of panchromatic images for the Pansharpening task.