RetinexDet: Enhancing Multispectral Object Detection via Retinex State Space Duality and Wavelet-Based Frequency Adaptive Fusion

Multispectral object detection has gained significant attention for its ability to enhance detection performance by integrating complementary information from both visible and infrared images. This approach has proven particularly beneficial across a wide range of industrial applications. However, the large number of parameters in existing network models and the dynamically changing features between dual modalities often result in inefficiencies, consuming excessive computational resources, and hindering effective fusion of the complementary data. This study introduces RetinexDet, a lightweight and efficient detection method designed specifically for infrared and visible image modalities. RetinexDet simultaneously addresses these two critical challenges by leveraging the Retinex state space duality block, which efficiently extracts and refines fine-grained features from both image types while emphasizing salient regions of interest. Furthermore, the Wavelet-Based Frequency Adaptive Fusion module is proposed, a novel method that adaptively fuses multilevel semantic and intensity information from the infrared and visible images in the frequency domain. Extensive experimental evaluations on multiple public datasets demonstrate the superior performance of RetinexDet. Specifically, on the Low-Light Visible and Infrared Paired dataset, RetinexDet reduces the model parameters by 98.8% compared to CrossFormer while achieving a higher mean average precision (mAP). When compared to GAFF, a method with a similar parameter count, RetinexDet outperforms with a 26.7% increase in mAP.