DFLMF-ISTD: Infrared small object detection network based on decoupled feature learning and multi-scale feature fusion

Abstract

Infrared small object detection is widely used in small maneuvering object awareness and high threatening object detection and recognition. In recent years, the introduction of deep learning methods has greatly improved the detection performance of infrared small objects. However, the presence of clutter in infrared small object images (low signal-to-noise ratio, SNR) and the lack of shape and texture information for the objects lead to a decrease in detection performance in complex environments. As such, in this article, an infrared small object detection network based on decoupled feature learning and multi-scale feature fusion is proposed. First, utilizing disentangled feature learning, we construct Reversible Column Networks (Revcol) with C3 modules to get RevcolC3 to alleviate the issues of complex feature extraction and the loss of small-scale object information. Second, a new lighted attention spatial pyramid pooling (LASP) module is proposed. By convolving the features extracted from the backbone and performing two consecutive pooling operations, a large kernel separated attention (LSKA) mechanism is introduced to process spatial and channel information separately. This enhances the model’s multi-feature extraction capabilities while reducing computational complexity. Finally, a novel lightweight three-dimensional multi-scale feature fusion (LTDMF) module is designed to efficiently utilize the correlations between three-level pyramid feature maps and effectively extract infrared small object features. This enhances the network’s ability to detect objects while maintaining the same model size. The proposed methodology is rigorously evaluated for its feasibility and reliability on the benchmark SIRST and IRSTD-1k datasets. The experimental results indicate that the proposed methodology outperforms current state-of-the-art (SOTA) infrared small object detection techniques under conditions of complex environments, small infrared object scales, and the absence of discernible texture and shape features.