Contextual Feature Modulation Network for Efficient Super-Resolution

Wandi Zhang, Hao Shen, Biao Zhang, Weidong Tian, Zhong-Qiu Zhao

Abstract: In recent years, single image super-resolution (SISR) reconstruction models based on convolutional neural networks (CNNs) have shown remarkable visual effects and reconstruction accuracy. However, the abundance number of parameters and relatively slow execution speed make it challenging to deploy these models on devices with limited memory and processing power. To address this challenge, we propose a Contextual Feature Modulation Network, denoted as CFMN, for efficient SISR tasks in this paper. This model successfully reduces model size and computational burden while maintaining high-quality image reconstruction. The proposed CFMN consists of a Multi-scale Feature Spatial Modulation (MFSM) and a Channel Attention Fusion Module (CAFM). Specifically, the MFSM replaces the traditional attention mechanism with a spatial modulation strategy. This module adaptively selects contextual feature representations at various scales and granularities through a multi-scale mechanism and a gated matrix, modulating the input features in the spatial dimension. Another core module CAFM complementarily extracts local contextual information and incorporates the Squeeze-and-Excitation Block to capture inter-channel dependencies. It effectively combines features from various channels through feature fusion, enhancing the network’s ability to perceive image details. The performance analysis demonstrates that the proposed CFMN effectively balances model complexity and performance.