| In recent years,object detection technology under normal illumination has gained significant development and many advanced detection algorithms have emerged.Object detection,as one of the important applications of computer vision,has been widely used in a large number of practical scenarios and also provides important support for the security field.For example,technologies such as face recognition are used in payment systems to provide security for commercial and financial data;intelligent surveillance systems identify illegal acts and dangerous objects to support the safety of public life and property.While image object detection in normal illumination levels currently yields sufficiently reliable results,these detectors do not perform as well in low illumination environments.Lowillumination environments are an almost inevitable scenario for various detection tasks,which pose a risk and threat to security in many situations,such as night-time face recognition,night-time behavioural surveillance,and detection of potholes in dimly lit environments.In low-light environments,the acquired images suffer from image degradation such as low visibility,low contrast,noise and colour bias,which seriously affects the extraction and recognition of key features and makes it difficult to produce highquality datasets.One of the major reasons for the scarcity of research results on low-light object detection is the lack of object detection datasets in low-light environments.In this paper,we address this phenomenon by investigating the topic from two aspects:unsupervised domain adaptive and lightweight networks,and propose the following two methods for low-light object detection.To address the problem of scarcity of datasets in low-illumination object detection,an unsupervised and adaptive object detection method in low-illumination environment is proposed.First,the pixel-level alignment module is proposed to perform image translation on the source domain images,which initially reduces the distribution difference between the source and target domains.Then,a multilayer feature alignment module combining noise attention is proposed to specifically deal with the problem of the presence of large amount of noise in low-illumination images,and an adversarial classifier and a gradient inversion layer are set in the multilayer network to promote further alignment of the source and target domains in terms of global features.Finally,an instance-level feature alignment module that fuses global features is proposed to compensate for the poor generalization of the instancelevel feature domain by introducing global information in the feature extraction network,and the features in the foreground region of the image are aligned to achieve unsupervised object detection.The experimental results show that the proposed method achieves better object detection for low-illumination images without using low-illumination dataset annotation and low-illumination image pre-processing.To address the problems of the lack of common roadbed obstacle datasets in low illumination and the large number of parameters and computation of existing object detection models,this paper produces the Dark Common Roadbed Obstacle(DCRO)dataset and proposes a light-weight model for Dark Common Roadbed Obstacle detection based on YOLOv5(LDRO-YOLOv5).Lightweight model for Dark Roadbed Obstacle detection based on YOLOv5).First,road images with normal illumination are collected and labeled with road obstacles,and the image style is adjusted to obtain a dataset of common obstacles on dark roadbeds.Subsequently,the MN＿CA module is proposed to improve the backbone network structure of YOLOv5 in order to lighten the model.The localisation loss function in YOLOv5 is improved to calculate the model localisation loss more accurately.Finally,the impact of the category imbalance problem in the dataset is attenuated using a label smoothing strategy,and the inference of the model is speeded up by structural reparameterisation.Through experimental validation on the DCRO dataset,LDROYOLOv5 achieves considerable results by improving the model detection accuracy while reducing the computational and parametric quantities of the model. |
