Research On Object Enhancement-Detection Method In Adverse Lighting Conditions

As one of the most challenging and fundamental tasks in the field of computer vision,object detection has a wide range of applications in practical production and life areas such as security monitoring,autonomous driving,and special operations.Existing object detection methods have achieved satisfactory performance on comprehensive datasets under large-scale normal lighting conditions.However,for images captured under adverse lighting conditions,the diversity of degradation in the input images,such as high contrast,texture loss,excessive noise,and domain shift,poses severe challenges to current object detection algorithms and restricts their application in complex scenes under all weather conditions.Therefore,theoretical research on image enhancement and object detection in complex lighting scenarios has significant practical significance and application value.To address this challenging problem,the mainstream approach is to use low-level methods(such as image enhancement)to pre-enhance the input images before the detection module,which is referred to as the two-stage enhancement-detection framework in this paper.These frameworks are mostly based on human visual appearance to enhance images captured in adverse lighting conditions,such as CNNbased supervised methods and GAN-based unsupervised methods.However,due to the inherent differences between human vision and computer vision,the evaluation metrics,network structures,and loss functions designed based on low-level tasks cannot adapt well to advanced visual tasks such as object detection.In this paper,a single-stage enhancement-detection framework is proposed to address the above problems,and the research content is as follows:(1)To address the efficiency issue of the two-stage enhancement-detection framework in complex lighting conditions,this paper proposes a single-stage enhancement-detection framework,ISPNet,which estimates the parameters of multiple ISP modules in the enhancement framework solely through the loss function of the detection task.ISP can perform various algorithmic processes on images,such as denoising,white balance,color correction,sharpening,color space transformation,and dynamic range enhancement,to improve image quality.This paper proposes and reorganizes differentiable ISP modules,enabling the learning of enhancement parameters through the detection loss and the combination of local-global network endto-end.Experimental results demonstrate that compared to the two-stage enhancementdetection framework,our method achieves performance improvement.(2)To address the difference in human-machine understanding between image enhancement modules designed based on human vision and object detection methods based on machine vision,this paper redefines the enhancement task as a parameter estimation task for the curve-Light Curve-that conforms to physical priors.First,the curve equation is heuristically derived based on the dual-range prior and atmospheric scattering model,and Light Curve is used to gradually enhance the input image.Then,a lightweight network is proposed based on Light Curve to learn the pixel-level curve parameter map.On this basis,a single-stage enhancement-detection framework,Light Net,is summarized and proposed for end-to-end object detection tasks in different lighting conditions.(3)This paper quantitatively tested the above methods on the Ex Dark and RTTS datasets for their improvement in object detection performance,and qualitatively analyzed the reasons for performance improvement from an interpretability perspective.Subsequently,ablation experiments were conducted to demonstrate the rationality of the algorithm design in this paper.Compared with the two-stage enhancement-detection framework designed based on human vision and the single-stage enhancementdetection framework ISPNet,the proposed Light Net achieved better results.Finally,the proposed method was deployed on low-power hardware platforms represented by FPGA to further verify its feasibility.