Abstract
Existing single image super-resolution (SISR) methods face the balance problem of reconstruction accuracy and computational cost caused by deep networks. As one of effective ways to realize network’s lightweight, distillation networks have attracted more attentions. However, they focus on improving reconstruction accuracy by extracting local feature information or global feature information individually. In order to address this issue, a novel lightweight Enhanced Self-modulated Feature Aggregation Distillation Network (ESFADNet) is proposed by fusing the idea of variance modulation, multi-scale, edge enhancement, and attention mechanism into distillation network. Different from traditional distillation networks, our network designs a new feature refinement module, named Efficient Feature Modulated Attention (EFMA), to enhance reconstruction accuracy by activating the interaction mechanism between local feature correlations and non-local contextual dependencies synergistically. Its key block, Efficient Self-Modulation Feature Aggregation (ESMFA), contains two parallel branches: an Edge-Enhanced Approximation of Self-Attention (EEASA) branch for capturing non-local feature information and a Multi-Scale Local Estimation (MSLE) branch to obtain the local feature information. And the developed PCMN is used to further optimize the representative features extracted from ESMFA in the spatial and channel dimensions. Extensive experiments illustrate that our proposed method achieves superior performances while with less computational cost compared to some state-of-the-art SISR methods.