Light Weight IBP Deep Residual Network for Image Super Resolution

Single-image super resolution (SR) is used to reconstruct a high-resolution image with more high-frequency details based on a low-resolution image as input. In recent years, image SR reconstruction based on deep learning methods has shown a considerably better performance than traditional methods. Early deep-learning-based methods deepen convolutional layers and directly reconstruct high-resolution images with complex neural networks. However, with the stacking of modules, network depth and model parameters increase, thereby raising computational resource; hence, it is difficult to apply on low-configuration devices. Furthermore, existing methods ignore the high-frequency details of the image, resulting in unsatisfactory performance. To solve these problems, a lightweight network model that applies the iterative back projection (IBP) mechanism to the network and reduces the dimensionality of the input image features is proposed. The proposed network model consists of three parts, namely, entrance module, main body module, and exit module. It designs a lightweight modular design to reduce the model calculation and control the network depth more easily by adjusting the number of ADB modules. The main part of the model consists of four lightweight accelerating deep residual back projection (ADB) modules. Each ADB module initially decreases the dimensionality of the input image features through a 1x1 convolutional layer to reduce the amount of calculation. Then, IBP is used to back project the image iteratively. Each ADB module only performs the downsampling back projection operation because the image features become larger after upsampling. Through three iterative downsampling back projection units, the high-frequency features of the image are fully explored, and the output image features are then compared with the shallow layer. Image feature fusion is used as input of the next ADB module, and the output part combines the high- and low-frequency image feature output by multiple ADB modules to complete the upsampling by using the PixelShuffle method to generate high-resolution images. Several experiments confirm that the proposed algorithm achieves better SR reconstruction accuracy with faster reconstruction speed than existing image SR methods.