Single Image Super Resolution Using Multi-path Convolutional Neural Network
Sedighe Dargahi, Ali Aghagolzadeh, Mehdi Ezoji · 2021
Nowadays, deep learning-based methods are used in single image super resolution tasks. Even though these methods yield acceptable results, they have the problems of requiring more data and memory space. In this paper, a method based on convolutional neural network is proposed, which not only requires less data and memory space for learning, but also leads to accurate results. In the proposed approach, a three-path network is presented which has the same subnetworks fed by different Low Resolution (LR) images. The original LR data are fed into the first path and its high frequency-discarded versions are fed into the other paths. In this way, the network is forced to detect and learn high level features from low information inputs via parallel feature extractors. Furthermore, multi-scale manner is utilized to simultaneously detect coarse and fine features. In addition, feature maps have been fully exploited by concatenating and merging them throughout the network. Moreover, a strategy of parallelizing upsampling methods is proposed to ensure the reconstruction of the High Resolution (HR) image is accurately verified. Experimental results demonstrate that the proposed network outperforms most of the state-of-the-art methods in terms of the depth, the number of parameters and input data, and overall complexity of the network. The proposed method also improves visual performance and reaches higher evaluation metrics compared to the most of the existing methods.