Research On Low Illumination Object Detection Method Baseed On Attention Mechanism And Domain Adaptation

Object detection is the most basic task in the field of computer vision,and it is also the basis and premise for completing more advanced visual tasks.It is widely used in security monitoring,intelligent assisted driving,and drone reconnaissance.Currently,deep learningbased object detection methods have achieved great success on normal illumination images,but there are few research results for low illumination object detection.In the case of insufficient lighting conditions,the acquired low illumination images have problems such as low contrast,loss of detail information,blurring and much noise.These problems lead to the direct use of existing object detection methods,resulting in low detection accuracy and inaccurate positioning of bounding boxes,which severely limits the development of technologies such as night monitoring and nighttime assisted driving.At the same time,due to the characteristics of low illumination images,capturing low illumination images and performing object-level annotation is more time-consuming and labor-intensive,and it is more difficult to build a large-scale low illumination object detection dataset.Aiming at the subject of low illumination object detection,research has been carried out from two directions: supervised learning and unsupervised domain adaptation.This dissertation proposes the following two low illumination object detection methods:1)Aiming at the problem that the existing object detection methods cannot achieve ideal results in low illumination scenes,a low illumination object detection method called Dark-YOLO was proposed from the direction of supervised learning,combined with multiscale features and attention mechanism,and improved detection head.Firstly,the feature extraction of low illumination images is carried out by using CSPDark Net-53 backbone network.Secondly,the path aggregation enhanced module is proposed to further enhance the ability of feature representation.Thirdly,the pyramid balanced attention module is designed to capture multi-scale features and effectively use them to generate features with different scales and more discriminative power.Finally,the Io U prediction branch is added to predict the Io U value for each prediction box to make the object location more accurate.The experimental results show that the proposed Dark-YOLO method has higher detection precision and more accurate bounding box localization in low illumination scenes compared with the current general object detection methods.2)Aiming at the problems of domain difference in object detection task and difficulty in making low illumination object detection dataset in low illumination scenes,a method of LIDA-YOLO(Low Illumination Domain Adaptation-You Only Look Once)is proposed based on the existing object detection dataset and YOLOv3 object detection network combined with unsupervised Domain Adaptation.This method solves the problem of domain difference from both local and global aspects.On the one hand,multi-scale local feature alignment is performed to reduce the low-level feature difference by obtaining three feature maps of different scales in Dark Net-53 network.On the other hand,multi-scale global feature alignment is performed on the three feature maps with different scales after Feature Pyramid Network(FPN)fusion to align the overall attributes of the image.Finally,the LIDA-YOLO method is compared with the most advanced unsupervised adaptive object detection methods on normal illumination PASCAL VOC dataset and low illumination Ex Dark dataset.Experimental results show that the LIDA-YOLO method outperforms stateof-the-art unsupervised adaptive object detection methods.