MPCNet: Multiscale Perception and Cross-Attention Feature Fusion Network for Infrared Small Target Detection

Infrared small target detection (IRSTD) is a challenging task due to the small size and low contrast of targets in infrared images. Traditional methods rely on extensive prior knowledge and exhibit poor robustness against complex backgrounds. Although CNN-based methods have made remarkable progress in IRSTD, effectively extracting and fully utilizing features at different levels remains challenging. To address this challenge, we propose two strategies for IRSTD: 1) multi-scale perception and 2) cross-attention feature fusion. Based on these strategies, a multi-scale perception and cross-attention feature fusion network is constructed, named MPCNet. Specifically, a multi-scale perception module in the encoder is designed to capture rich contextual information and enhance the localization perception of small targets. For cross-attention feature fusion, a global semantic-aware fusion module in the encoder and a semantic-guided cross-attention fusion module in the decoder is devised, respectively. The former achieves more refined feature fusion by narrowing the semantic gap between features at different levels, while the latter further captures target features accurately by enhancing the semantic associations between features in the encoder and decoder. Experimental results verify the effectiveness of the proposed MPCNet compared to other state-of-the-art (SOTA) IRSTD methods in both quantitative and qualitative evaluations. The code will be released on https://github.com/Wangtao-Bao/MPCNet.