Deep Networks for Image Super-Resolution with Sparse Prior

Abstract

Deep learning techniques have been successfully applied in many areas of computer vision, including low-level image restoration problems. For image super-resolution, several models based on deep neural networks have been recently proposed and attained superior performance that overshadows all previous handcrafted models. The question then arises whether large-capacity and data-driven models have become the dominant solution to the ill-posed superresolution problem. In this paper, we argue that domain expertise represented by the conventional sparse coding model is still valuable, and it can be combined with the key ingredients of deep learning to achieve further improved results. We show that a sparse coding model particularly designed for super-resolution can be incarnated as a neural network, and trained in a cascaded structure from end to end. The interpretation of the network based on sparse coding leads to much more efficient and effective training, as well as a reduced model size. Our model is evaluated on a wide range of images, and shows clear advantage over existing state-of-the-art methods in terms of both restoration accuracy and human subjective quality.

Superresolution Reconstruction of Electrical Equipment Incipient Fault

“…The sparse code is multiplied with HR dictionary D x in the last linear layer to reconstruct HR patch. This stage is Figure 4: Process of middle mapping stage S 2 [22].…”

“…As a result, HR version of the incipient fault on electrical equipment can be more visible and helpful in achieving easier fault detection. The recurrent middle stage adopted the idea referred in [22], which is mathematized as Eq. (2) where y is the input signal and h is an coordinatewise shrinkage function defined as…”

“…In this paper, we extract 100-dimension feature in this stage. The middle stage of this model S 2 is a high-dimension mapping process, which includes three linear layers shown in Figure 4 [22]. The sparse code is multiplied with HR dictionary D x in the last linear layer to reconstruct HR patch.…”

“…The matrices W and S are learned so as to minimize the approximation error to the optimal sparse coding on a given dataset. The final convolution stage S 3 producing HR reconstruction result is described as S 3 (y) = w 3 * S 2 (y)+b 3 , b 3 is a vector, and w 3 project the coefficients to image spatial domain to get averaged HR result [22,23].…”

Single Image Super Resolution : A Comparative Study

“…If the images haven"t insufficient patch self-similarity, these methods are not able to produce satisfying results [9]. A recent methods proposed in [17] moderate this limitation by learning image prior models via kernel principal component analysis from multiple image frames.…”