HMANet: Hierarchical Multi-Attention Enhancement Network for infrared small target detection

Abstract

The application of infrared small target detection technology is becoming increasingly widespread in both military and civilian fields, with its strategic value and practical significance being continuously emphasized. Existing CNN-based methods are highly prone to the loss of critical target features in pooling layers and exhibit limited capability in capturing local features of small targets. This leads to significant degradation of spatial information, making it challenging to effectively differentiate small targets from background clutter. As a result, such methods are not directly suitable for infrared small target detection. To address the aforementioned issues, we propose a Hierarchical Multi-Attention Enhancement Network (HMANet) tailored for the precise localization of small infrared targets. Firstly, we design a fusion strategy that integrates spatial-channel attention with frequency-domain attention. This approach builds inter-layer feature correlations using channel attention in the spatial domain. It enhances fine-grained details through multi-scale spectral processing in the frequency domain. Cross-domain feature interaction further improves target-to-background contrast in a hierarchical manner. Secondly, we capture contextual dependencies across feature layers by modeling the relationships between embedded tokens at multiple levels using diverse attention interaction modules. This enables the construction of hierarchical contextual representations, facilitating the accurate detection of infrared small targets. In addition, comprehensive evaluations on the IRSTD-1K and SIRST3 dataset demonstrate that the proposed method achieves competitive performance across all evaluation metrics.