Abstract: Modern deep learning techniques outperform cutting-edge signal processing techniques in image restoration applications. However, these CNNS still perform poorly if a picture has a lot of patterns and structures. Here, we offer a unique feature space deep residuary learning technique that outperforms the current residual learning method to handle this problem. The main concept is derived from the finding that a learning algorithm performs better if the input and/or label manifolds may be topologically simplified by an analytical planning to a feature space. Our in-depth numerical analyses utilizing denoising tests and the single-image super-resolution (SISR) rivalry show that proposed space residual learning surpasses the present atatus of the workmanship methods. Furthermore, our algorithm placed third in the NTIRE competition with a computational time that was 5–10 times quicker than that of the top-placed teams