“…It can be seen that compared with the classical faster R-CNN, our algorithm has higher detection accuracy, and compared with RetinaNet, YOLOv3, and CenterNet, it still has fewer FLOPs and higher detection accuracy. Compared with normalO2-DNet, DDQ-DETR, and ARSD, 46 CSnNet, 47 our algorithm not only has better detection accuracy but also dramatically reduces the number of parameters, mainly because the backbone network depth of the normalO2-DNet is too deep, which leads to a large number of model parameters, and the model ignores the false detection and missed detection of small objects; VDNET-RSI constructs a two-stage object detection algorithm combined with super-resolution reconstruction, which improves the resolution of input image, but increases the parameters of the model; DDQ-DETR has more dense query keys, which improves the multi-scale features and increases the computation of the model. The knowledge distillation network designed in ARSD algorithm ensures the lightweight of the model, but it is difficult to fundamentally improve the accuracy of small object detection only by using multi-scale feature fusion.…”