A Multi-level Wavelet Decomposition Network for Image Super Resolution

Machine learning with deep convolutional neural network has seen huge adaptation in computer vision applications over the last decade. In this paper, we propose a novel network architecture to perform single image super resolution based on deep convolutional neural network (DCNN) and discrete wavelet transform (DWT). In fact, the discrete wavelet transform is applied in multi-levels on the low resolution image to divide it into four sub-bands. Then, the deep convolutional neural networks is applied only on the approximation sub-band of second level in order to reduce computational requirement. Also, we discovered that dropping the DD sub-band does not impact the perceptual quality of the reconstructed image. So we propose to replace by a matrix of zero. By training various images such as Set5 and Set14 datasets, good results are obtained allowing to validate the effectiveness and efficiency of the proposed method based on evaluation of PSNR and SSIM. The reconstructed image achieves a high resolution value in less run time than existing methods.