（1.School of Mathematics and Statistics， Changsha University of Science and Technology， Changsha 410114， China；2.School of Computational Science and Electronics， Hunan Institute of Engineering， Xiangtan 411104，China） 在知网中查找 在百度中查找 在本站中查找
To solve the problem of large computation and low accuracy of the current mainstream algorithms for small object detection， this paper replaces the backbone network in YOLOv4 with the lightweight network MobileNetV3， and replaces some ordinary convolutions in the neck network with depthwise separable convolutions. At the same time， a new loss function IF-EIoU Loss is defined for small object detection. Therefore， MDS-YOLO object detection model is constructed. This model has a high detection speed and good detection performance for small object. To verify the effectiveness of the model， experiments are carried out on MS COCO dataset and Visdrone2019 dataset， respectively. Compared with the YOLOv4 algorithm， on MS COCO dataset， the average detection accuracy of the MDS-YOLO algorithm is improved by 1.5 percentage points， the detection accuracy of small object is increased by 3.3 percentage points， and the detection speed is also increased from 31 frames per second to 36 frames per second. On the Visdrone2019 dataset， the MDS-YOLO algorithm increases the average detection accuracy from 14.9% of YOLOv4 to 16.3%. The experimental results show that the MDS-YOLO algorithm proposed can effectively improve the detection accuracy of small object.
