| Object detection is one of the most basic tasks in the field of computer vision.It is widely used in practical applications such as industrial detection,autonomous driving,and urban monitoring.Compared with normal-light images,low-light images have blurred content,low contrast,high noise and loss of details.It is difficult for the object detector to extract useful features,which affects the detection performance of the object detector.Moreover,due to the inconsistent distribution of the normal-light image dataset and the lowlight image dataset,there are significant differences in the features extracted by the object detection model from the two sets of datasets.Therefore,the model trained using normallight images cannot effectively detect objects in low-light images.In addition,the existing object detection algorithms rely heavily on labeled datasets.However,there are few publicly available low-light image datasets,and manually making datasets is time-consuming and laborious.The thesis investigates object detection in low-light conditions by studying techniques such as feature optimization and unsupervised domain adaptation.It addresses the problem of reduced detection performance in low-light images due to insufficient information in image features,scarce samples,and domain shift.The proposed approach improves the detection performance of object detectors on low-light images.The thesis proposes the following two low-light image detection methods:Due to the low contrast,severe detail loss,and high noise of low-light images,existing object detection algorithms perform poorly on these images.To solve this problem,a lowlight object detection method that combines spatial-aware attention mechanism with multiscale feature fusion is proposed.Firstly,multi-scale features are fused by multi-scale interactive memory pyramid to enhance effective information under low-light condition,and features of memory vector storage samples are set to capture potential correlation between samples.Then,a spatial-aware attention mechanism is introduced to obtain long-distance context information and local information of features in spatial domain,thereby enhancing the object features in low-light images.Finally,multiple receptive field enhancement module is used to expand receptive field of the features,allowing the detection network to dynamically adjust the receptive field size based on the multi-scale feature information.Experimental results on the Ex Dark dataset show that m AP of the proposed method reaches77.04%,which is 2.6%～14.34% higher than existing mainstream object detection methods.Aiming at the lack of low-light image datasets and the degradation of object detector detection performance caused by domain shift,the thesis proposes a low-light object detection method based on multi-scale unsupervised domain adaptation.This method transfers the knowledge learned from the labeled normal-light image dataset to the unlabeled low-light image dataset,in order to achieve more accurate cross-domain target detection.Firstly,the pixel-level and feature-level domain classifiers are introduced on the three different scale features output by the object detector to reduce the distribution difference between the source domain and the target domain at the pixel level and the feature level.Then,a dynamic category alignment module is proposed,which takes into account the different sample distributions in the domain and class.It assigns greater weights to categories with scarce samples and samples that are easy to adapt,effectively compensating for the difference in data distribution.Finally,through the global-local alignment module,the feature gradually transition from global alignment to local alignment during the training process,so that the domain adaptation network focuses on the foreground region that is crucial for object detection.Experiments show that m AP of the proposed method reaches63.78 %,which is 10.45 % ～ 15.98 % higher than the existing cross-domain detection model. |
