Lightweight infrared image super-resolution reconstruction network with contrast-driven self-modulation aggregation

Abstract

With the widespread application of infrared imaging technology in various fields, the demand for infrared image resolution is constantly increasing. However, the resolution of infrared images is limited by imaging hardware and environmental conditions. Additionally, current state-of-the-art image super-resolution methods predominantly target visible light imagery, depend on deep network architectures, and demand substantial computational resources and high-end hardware. Therefore, we propose a lightweight infrared image super-resolution reconstruction method based on the contrast-driven self-modulation aggregation network (CDSMANet). Firstly, a core module is designed to decompose infrared images into high-frequency, medium-frequency, and low-frequency components and achieve fusion interactions to drive the extraction of both local and non-local features. It performs a more accurate reconstruction through self-modulation aggregation. Specifically, we generate three branches of different frequency features through feature separation. The medium-frequency and low-frequency branches extract non-local features through non-local self-attention approximation, while the high-frequency branch models local information and extracts local details. Secondly, an adaptive multi-receptive field fusion module (AMF) is developed to integrate these different features, enabling mutual driving of feature extraction. Moreover, a multi-scale convolutional pooling feedforward network (MCPN) is used to further capture deep importance features. Experiments have shown that CDSMANet achieves a good balance between reconstruction performance and computational efficiency on public infrared image datasets.