Research On Key Technologies Of Object Detection In Low-light Scenes

As the most basic and challenging task in computer vision,object detection has a wide range of applications in practical fields such as security monitoring and pedestrian detection.The diversity of low-light environments brings severe challenges to various object detection algorithms and restricts their application in complex scenes such as allweather traffic monitoring and autonomous driving.Therefore,the research on related theories and technologies such as image enhancement and object detection in low-light scenes has practical significance and application value.This paper starts with low-light image enhancement and object detection,focuses on the object detection algorithm in low-light scenes,and carries out exploration and research.On the one hand,this paper proposes a low-light image enhancement algorithm based on a generative confrontation network,which explores image enhancement methods in real low-light scenes lacking paired data;on the other hand,this paper proposes an end-to-end low-light image object detection algorithm based on the dynamic enhancement network,and studies how to improve the performance of the object detection algorithm in low-light scenes.The specific work of this paper is as follows:1.Aiming at image quality problems such as noise,low brightness and low contrast in low-light scenes,a low-light image enhancement method based on a generative confrontation network is proposed.High-quality images can be recovered from low-light images without paired data.First,by constructing exposure and noise reference images as additional inputs to the generator,the generator is guided to focus on the underexposed and noisy areas in the image.Second,this paper introduces a global self-attention module in the generator,and uses the self-attention mechanism to model the global semantic representation of the image,so that the generator can pay attention to the global information when enhancing the low-light image.At the same time,the content consistency and smoothness of the enhanced image are constrained by adding perceptual loss and total variation loss to the generator.Finally,this paper proposes a multi-scale discriminator based on a full convolutional network to help the discriminator to efficiently capture the global and local features of the input image,and guide the generator to generate enhanced images with good overall quality and local details.The experimental results on the standard test images of the public dataset show that the method proposed in this paper can effectively generate high-quality images with good visual perception.2.In order to solve the problem that the existing low-light image enhancement methods have limited improvement in object detection performance,an end-to-end lowlight image object detection method is proposed.In the enhancement stage,this article uses several enhancement subnetworks to simulate a set of enhancement methods,and on this basis,further suppress noise and improve contrast.The enhancement subnetwork consists of two parts.The dynamic filter network is used to generate a sample-specific convolution kernel to filter the input image,and the adaptive exposure module generates the exposure map of the corresponding image to enhance the dark details of the image.The enhancement stage also gives a set of weights corresponding to the enhancement subnetworks,which indicate the degree of influence of the corresponding subnetwork on the detection result.The detection stage is based on the classic two-stage object detection algorithm,which generates high-quality region proposals through the weighted average of the classification loss of the region proposal network,thereby improving the overall performance of the detector.This paper conducts verification experiments on the public low-light object detection dataset,and the results show that the method in this paper can effectively improve the detection performance.