Research On Lightweight Super-resolution Convolutional Neural Networks

In the case of limitations of hardware devices,super-resolution algorithm is used to improve the resolution of the image,recover the details of image and reconstruct high-quality image.Deep learning method based on convolutional neural network can effectively extract the internal features of the image,learn the mapping relationship between the low-resolution image and the high-resolution image to reconstruct high-resolution image.Based on the convolutional neural network,this paper focuses on the efficiency of the super-resolution network.The main works are summarized as follows:At present,the super-resolution network model is relatively deep,and demands large amount of parameters and computational resources,so it is hard to be applied to resolve the real-world problems.Based on the convolutional neural network,this paper concentrates on technologies of lightweight convolutional neural networks,designs a lightweight,fast and effective super-resolution network.In this network,the group convolution is firstly used to reduce model parameters effectively,and 1×1convolution layer is employed to create a linear combination of the output of the group convolution layer.Meanwhile,feature reuse is encouraged through dense connection to improve network efficiency.In order to further improve the quality of reconstruction image,the attention mechanism is developed to boost the feature discriminative ability of the network.Experimental results show that the proposed method not only recovers image details effectively,but also has faster running speed and fewer parameters.The multi-scale information of image is insufficiently utilized by super-resolution network,and the information of different receptive fields is linearly aggregated,so the size of receptive fields can not be adaptively changed.In order to fully extract the different scale features of images,a multi-scale residual super-resolution network based on dynamic selection mechanism of convolutional kernel is proposed.Firstly,the selective kernel convolution is presented to extract multi-scale information from feature maps of different scales.On this basis,the residual nested structure is designed to make full use of the local feature information and strengthen the information flow during back propagation.In addition,the weight normalization layer is introduced to normalize the weight of each layer to accelerate network convergence and improve the performance of the network.Experimental results show the proposed method achieves better results on both objective and subjective evaluation indicators compared with the state-of-the-art super-resolution methods,and only contains a few parameters.