Image Super Resolution Based on Fully Convolutional Multi-path Model with Adaptive Skip Connection
Zhixin Zeng, Hong Liu, Shengchun Wang · 2019
Notable success is achieved by deep learning based method in the field of single image super-resolution(SISR). However, the monotonous chain-like design of neural network is widely used among those achievement, which the model could not adequately utilizing the characteristic of different neural network architecture. In this paper, we propose SISR based on a learning-based multi-path with adaptive skip connection model in order to enhance the ramification of super-resolution reconstruction and to fully explore the characteristic of diverse neural network architecture. The proposed method utilizes the characteristics that is coming from different architecture to extract the high-frequency and low-frequency feature of the original image separately through the deep and shallow path. The adaptive skip connection is deployed in the deep path to boost the correlation of the feature extracted which is coming from different layers. The last part that the fusion and reconstruction module use the features which were extracted by previously different path to construct the high-resolution version of the origin image. Urban100 is adopted to train proposed model. Set5, Set14 and BSD100 are utilized to test our model trained. Experimental result verify that the proposed method surmounts the other comparative method and perform well.