An Improved Algorithm For Object Detection Based On YOLO Series

The task of object detection is an important subject in the field of computer vision and has a wide range of practical applications in industry.In this paper,we focus on the task of object detection,and propose an improved object detection algorithm based on YOLOv5.In addition,we conduct experiments on the challenging MS COCO2017 dataset.The experimental result is that based on YOLOv5 s and YOLOv5 m,our model improves the accuracy of the detection especially for medium and small targets,while basically maintaining the inference speed.The main contents of this paper can be divided into the following parts:The first part,we introduce the research background of object detection,summarize the typical object detection algorithms based on deep learning,and expound the typical methods of feature fusion.The second part,we introduce the latest algorithm of the YOLO series,YOLOv5.Specifically,we expound YOLOv5 algorithm from these aspects: data augmentation,network structure,loss function and so on.The third part,we propose a novel feature fusion method.Firstly,considering the advantages of BiFPN structure,we apply the bidirectional feature fusion method to our detection network and change the original fusion strategy of pixel-summation to channel concatenate.Secondly,we make full use of the information of low-level features in order to improve the detection ability of small targets.The experimental part,we select the challenging MS COCO2017 datasets for experiment which including 110,000 images.Compared with YOLOv5 s,the average precision of our model is improved from 36.7% to 38.1% and the recall is increased from57.4% to 58.4%.The average precision of our model is improved from 21.0% to 22.7%on the small scale objects,and from 42.1% to 43.0% on the medium scale objects;Compared with yolov5 m,the average precision of our improved model is increased from44.5% to 44.7%,and from 27.4% to 28.6% on the small scale objects.Moreover,the loss function is changed from binary cross-entropy loss to focal loss for comparative experiments.We find that focal loss does not necessarily improve the accuracy.Furthermore,we compare our improved algorithm with other similar methods.The results show that based on YOLOv5 s and YOLOv5 m,our model is improved on both detection average precision and recall,especially for small and medium scale targets.Finally,the speed measurement experiment shows that our model can not only improve the average precision but also basically maintain the speed of inference.