A Lightweight Network for Detecting Small Targets in the Air

Fast and accurate detection and identification of small airborne targets are of great importance, to security in the air. Unmanned aerial vehicle detection algorithms are mostly deployed on edge devices, and a yolov5-based aerial target lightweight detector is proposed by compressing channels and network cropping for the limited resource characteristics on edge devices. Firstly, the shallow cross-stage partial module is extended and optimized when designing the feature ex-traction network to maximize the use of shallow features. Secondly, the network is cropped to reduce the number of down-sampling, which makes the computation faster. Finally, the pyramid network used for feature fusion is simplified by modifying from two upsampling operations and two downsampling operations to only one upsampling operation. On the homemade dataset, the proposed Yolo-mini achieves 94.44% mean average accuracy on the test set and the Giga floating-point operations per second of the model is only 3.2, which achieves a better balance of accuracy and computation compared to other lightweight algorithms.