Low-light scene object detection with infrared sensing

ZHANG Zhijia; NA Xingqi; XIAO Yuhang; FANG Jian; ZHAO Huaici

doi:10.7688/j.issn.1000-1646.2025.04.02

Abstract：[Objective]With the rapid development of artificial intelligence,object detection technology based on visible light images has become increasingly advanced and has been widely applied in fields such as autonomous driving,security monitoring,and intelligent transportation.However,in low-light scenes(such as nighttime or dimly lit environments),the performance of object detection algorithms based on visible light images decreases significantly.This is primarily due to severe information loss in visible light images under low-light conditions,making it difficult to extract target features.To solve this problem,multi-modal object detection technology combining visible light and infrared images was proposed,which could effectively enhance object detection performance in low-light scenes.However,the multi-modal method is costly and requires precise registration of images from different modalities,which increases system complexity and processing burden.In response,this study proposed an object detection network with infrared sensing(InSCnet),aimed at using a visible light camera to predict infrared thermal radiation characteristics,thus improving the network's object detection capability in low-light scenes without increasing modality.[Methods]The InSCnet network used visible light images as input and generated infrared images through an infrared prediction branch(IPB),which predicted thermal radiation characteristics to enhance the network's perception of low-light scenes.A complementary fusion filter(COFF)module was designed to effectively integrate multi-scale visual and thermal radiation features.By complementing these two features,the COFF module enhanced their mutual complementarity and avoided the network's over-reliance on a single modality.In addition,a hybrid feature pyramid(HyFP)module was employed to further improve the fusion and extraction of multi-scale global and local features through feature pyramids and attention mechanisms,ensuring that the network maintained high detection accuracy under varying low-light conditions.[Results]Experimental results show that InSCnet performs excellently on the LLVIP pedestrian detection dataset,with SmAP50 reaching 0.830 and SmAP50-95 reaching 0.426.Moreover,experiments conducted on the DroneVehicle dataset show a SmAP50 of 0.702,confirming its ability to handle multi-class low-light detection.[Conclusion]InSCnet improves object detection performance in low-light scenes by introducing infrared thermal radiation characteristics and a feature fusion mechanism.The network can effectively detect objects that are difficult to identify in visible light images under low-light conditions,providing an effective solution for object detection in such environments.Future research will further explore ways to optimize the network structure.