LightYOLO-S: a lightweight algorithm for detecting small targets

Small target detection tasks are the focus and difficulty of target detection tasks. Methods to improve detection accuracy are often accompanied by drawbacks, such as a high number of parameters, computational effort, and latency. Therefore, this paper proposes the LightYOLO-S target detection algorithm based on YOLOv8s, which achieves high accuracy in small target detection tasks. First, the proposed LightC2f module reduces the overall number of parameters, computation, and inference time of the model while maintaining the same plug-and-play characteristics as the C2f module. Second, the proposed Wise-DIoU loss function, which speeds up model convergence and improves accuracy without increasing the number of parameters or computation. Third, the proposed Dynamic Sampler counts the IoU scores and classification scores of the screened training samples to adjust the sample allocation function, so that the model can obtain the best training samples during training.The results of experimental studies on the VisDrone2019 UAV aerial photography dataset and the DOTA remote sensing dataset indicate that LightYOLO-S exhibits greater accuracy and faster operation speed than the current state-of-the-art detection algorithms.