Learning lightweight super-resolution networks with weight pruning. (December 2021)

• Record Type: Journal Article
• Authors: Jiang, Xinrui; Wang, Nannan; Xin, Jingwei; Xia, Xiaobo; Yang, Xi; Gao, Xinbo
• Is Part Of: Neural networks. Volume 144(2021)
• Page Start: 21
• Page End: 32
• Publication Date: 2021-12
• Subjects: Image super-resolution -- Lightweight network -- Model pruning; Neural computers -- Periodicals; Neural networks (Computer science) -- Periodicals; Neural networks (Neurobiology) -- Periodicals; Nervous System -- Periodicals; Ordinateurs neuronaux -- Périodiques; Réseaux neuronaux (Informatique) -- Périodiques; Réseaux neuronaux (Neurobiologie) -- Périodiques; Neural computers; Neural networks (Computer science); Neural networks (Neurobiology); Periodicals; 006.32
• DOI: 10.1016/j.neunet.2021.08.002
• Languages: English
• ISSNs: 0893-6080

Abstract

Abstract: Single image super-resolution (SISR) has achieved significant performance improvements due to the deep convolutional neural networks (CNN). However, the deep learning-based method is computationally intensive and memory demanding, which limit its practical deployment, especially for mobile devices. Focusing on this issue, in this paper, we present a novel approach to compress SR networks by weight pruning. To achieve this goal, firstly, we explore a progressive optimization method to gradually zero out the redundant parameters. Then, we construct a sparse-aware attention module by exploring a pruning-based well-suited attention strategy. Finally, we propose an information multi-slicing network which extracts and integrates multi-scale features at a granular level to acquire a more lightweight and accurate SR network. Extensive experiments reflect the pruning method could reduce the model size without a noticeable drop in performance, making it possible to apply the start-of-the-art SR models in the real-world applications. Furthermore, our proposed pruning versions could achieve better accuracy and visual improvements than state-of-the-art methods. Highlights: We prove most existing SR networks are over-parameterized and the model size can be dramatically reduced without a noticeable drop in performance, making it possible to apply the SR model in real-world applications. We introduce a progressive global sparse optimization method to prune the SR network and explore a sparse-aware attention module to reduce the performance gap between the pruning version and the original one.