High-Speed Motion Target Real-Time Detection Based on Lightweight Deep Feature Learning Network

Real-time detection of high-speed moving targets is a significant and wildly applicable subject in the field of target detection. It is extremely important in both civilian and military fields. However, most of the current mainstream object detection models have difficulties in fulfilling requirements of accuracy and speed simultaneously. In this article, we propose a lightweight deep feature learning network (LDFL-Network), which can accurately complete detection of high-speed objects in real time. The network utilizes Ghost convolution instead of ordinary convolution to extract image features because of its superior performance in developing a lightweight and efficient network architecture. Utilizing the principles of Ghost convolution, we propose the I-GhostConv module, which employs the involution algorithm to enable flexible feature extraction and improves the network's expressive capacity. In addition, we propose a novel feature fusion method: adaptive cross-fusion structure. This structure adaptively allocates weights to feature maps and integrates various feature information crossly and comprehensively. Experiments are conducted on two representative datasets for high-speed motion target detection. The results of experiments demonstrate that the proposed network can accurately detect the targets in high-speed motion and achieve accuracy speed higher than mainstream target detection models.