Frequency-Aware and Conditional Entropy-Guided Network for Multispectral Object Detection

Abstract

Multispectral object detection is a technique that integrates information from different spectral bands, such as visible and infrared imagery, to improve the reliability and robustness of detection in complex environments. However, existing methods struggle to effectively integrate infrared structural contours and visible textural details. In addition, prevalent fusion strategies, especially those built upon cross-attention mechanisms, tend to prioritize modality-shared features while neglecting in-depth exploration of modality-specific characteristics. To tackle these issues, this article proposes a novel frequency-aware and conditional entropy-guided network. Specifically, a dual-domain wavelet enhancement module (DWEM) using the stationary wavelet transform is designed to capture translation-invariant multiscale features, thus preserving precise spatial structures crucial for accurate object localization. Furthermore, a conditional entropy-guided cross-attention fusion module, which leverages local conditional entropy to quantify cross-modal information gain, is developed to enhance the distinct features from each modality. Extensive experiments on three benchmark datasets demonstrate that the proposed method delivers competitive performance under diverse and challenging conditions, providing a robust and effective solution for real-world deployment across varying scenarios.