Improvement Of Small Object Location And Detection Algorithm Based On Non-fixed Anchor

The task of object detection is to identify and locate objects in optical images or videos.In other words,it is necessary to indicate the category of the object and the location of the object.Because object detection needs to be applied to life scenes or some more complex scenes,it is necessary to make the object detection algorithm real-time and accurate,that is,it can quickly detect the object and correctly locate and classify it.With the development of artificial intelligence technology,deep learning is used more and more widely in the field of object detection,and people's production and life also have object detection applications everywhere.For example,self-driving cars,face recognition systems,video surveillance videos,and the detection of diseased tissues in medical images,these devices will all use object detection,so the research of object detection algorithms is getting more and more attention.With the development of object detection technology progress,people's lives become more convenient and faster.The accuracy of small object detection is always the difficult problem in objection detection.Missing rate of small object and precise localization are two key factors that limit current small object detection systems.We propose a novel two-stage detection method,MLR-SODNet(Multipoint Local Regression-Small Object Detection Network),collectively addresses both small object missing and precise localization problem.For object missing,we propose a novel backbone,named SODNet(Small Object Detection Network),which design the novel convolutional stages to maintain high spatial resolution in deeper layers of traditional backbone network for image classification.For precise localization,we propose a MLR(Multipoint Local Regression)module to predict multiple bounding-box offsets and average them to get a final offset for an object proposal,besides,we add T-Io U(Twice-Intersection over Union)prediction branch to the overall network to calculate Io U twice at inference to eliminate the feature offsets of Ro Is to obtain more accurately localized bounding-box.Experiments performed on MS COCO demonstrate that MLR-SODNet achieves state-of-the-art detection accuracy with high efficiency.On MS COCO test-dev,our MLR-SODNet performs better than existing two-stage methods.The average accuracy rate of small objects(AP_S)increased by 1.9%compared with the original method,the average accuracy rate of medium objects(AP_M)increased by 3.6%,and the average accuracy rate of large objects(AP_L)increased by 5.8%.