Font Size: 

YOLOv2 + On-Category (OC) detection system combines YOLOv2 and On-Category Anchors technique to improve the accuracy for single-stage object detection, which is to create anchors based on the categories of foreground objects. However, it will result in the decrease of the accuracy when FPN was added to the system. To improve the accuracy of YOLOv2 + OC detection system, some improvements are given in this paper. The key ingredients points of our improvements are given as follows: 1) FPN (Feature Pyramid Network) is adjusted to suit YOLOv2 + OC detection system. Lower-level features are downsampled and a feature pyramid is established with multi-level features through lateral connections. So that the number of anchors can be reduced greatly, and YOLOv2 + OC detection system can be adopted to detect images by using their multi-level features. 2) A shared detection module is established for YOLOv2 + OC detection system. Feature maps of each layer from feature pyramids are used to train one detector only, which can make YOLOv2 + OC detection system more robust. 3) Mosaic data augmentation is applied in YOLOv2 + OC detection system. By fusing every four images into one image, the semantic information in a single image is increased and the data set is expanded. The results illustrate that the improvement in average precision (AP) is up to 3.6% and the detection speed is 33.3FPS, which could meet the requirements of real-time detection.