Research On Image Super-resolution Method Based On Multi-view Convolutional Neural Network

Deep networks can learn image features and more accurate information better than traditional super-resolution networks to express clean images without the influence of noise.They have made substantial progress in image super-resolution.However,most of these networks rely on deeper architectures to enhance the sharpness of predicted images,and a single feature cannot handle complex screens well in image super-resolution network models.(1)In the image super-resolution method based on a double convolutional neural network,a neural network with multi-branch network is proposed to obtain high-quality images.The neural network of multi-branch network relies on two sub-networks to extract complementary low-frequency features to enhance the learning ability of the network.To prevent the long-term dependency problem,a combination of convolutional and residual learning operations is embedded into the dual sub-network.To prevent the loss of information from the original image,enhancement blocks are used to collect the original information and obtain deeper high-frequency information through sub-pixel convolution.To obtain more high-frequency features,feature learning blocks are used to learn more high-frequency information details.Experimental results show that the neural network of the proposed multi-branch network outperforms other popular super-resolution networks.(2)In the image super-resolution method based on a deformed heterogeneous convolutional neural network,a parallel heterogeneous convolutional neural network for image super-resolution is proposed to solve the problem that the existing model is not enough to satisfy various image super-resolution Discrimination problem.The image super-resolution parallel heterogeneous convolutional neural network adopts the form of multi-view blocks,and uses two completely different model structures to super-resolution the same image to improve the learning ability of the network model.In terms of parameters,to reduce the model parameters,considering the limitations of the equipment,a lightweight network model is used to solve the feature extraction of images.Considering that due to the use of fewer parameters and model layers,to balance the loss of features caused by the reduction of parameters,a form of heterogeneous network is added to balance the regulation between parameters and the performance of the super-resolution model.Experiments show that the inpainted super-resolution images outperform most of the traditional super-resolution image models.