Small Object Detection for Drone Image Based on Advanced YOLOv7
Hao Chen, Jingyu Wang, Jingwei Li, Yu Qiu, Dengyin Zhang · 2023
Deep learning models for object detection on drone image have become a typical solution as consumer drone become more and more popular. Challenges in this scene, such as drone move too quick or too high, blocked by buildings or trees, make the accuracy of detection low. In response to the aforementioned issues, this paper selects YOLOv7 as a real-time object detector to improve, and designs the model's backbone network using a large convolution kernel architecture to expand the effective receptive field of convolution. Simultaneously, a BiFPN-like struct is designed for the YOLOv7 and used to process feature maps of various scales in order to solve the problem of the model's low precision in the drone scene. To support the large kernel backbone architecture, this paper uses methods such as struct reparameterization and depth-wise convolutions, so that the model has better inference performance.