Ensemble based deep networks for image super-resolution. (August 2017)

• Record Type: Journal Article
• Authors: Wang, Lingfeng; Huang, Zehao; Gong, Yongchao; Pan, Chunhong
• Is Part Of: Pattern recognition. Volume 68(2017:Aug.)
• Page Start: 191
• Page End: 198
• Publication Date: 2017-08
• Subjects: Super-resolution -- Ensemble -- Sparse prior -- Deep networks; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.02.027
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We suggest a general super-resolution method by using the ensemble of multiple models to reconstruct high-resolution results. We focus on Sparse Coding Network and propose a new state-of-the-art method named Ensemble based Sparse Coding Network. A back-projection method is introduced to solve the weights of different ensemble models. We accelerate the training of Sparse Coding Network (about 120x faster) by using batch normalization. Abstract: There have been significant advances in deep learning based single-image super-resolution (SISR) recently. With the advantage of deep neural networks, deep learning based methods can learn the mapping from low-resolution (LR) space to high-resolution (HR) space in an end-to-end manner. However, most of them only use a single model to generate HR result. This brings two drawbacks: (1) the risk of getting stuck in local optima and (2) the limited representational ability of single model when handling various input LR images. To overcome these problems, we novelly suggest a general way through introducing the idea of ensemble into SR task. Furthermore, instead of simple averaging, we propose a back-projection method to determine the weights of different models adaptively. In this paper, we focus on sparse coding network and propose ensemble based sparse coding network (ESCN). Through the combination of multiple models, our ESCN can generate more robust reconstructed results and achieve state-of-the-art performance.