Abstract: In order to solve the problem of high omission ratio of small target objects in remote sensing image object detection, this paper puts forward a small target detection method EfficientNet-YOLOv3 for remote sensing images requiring higher detection accuracy. The method, based on YOLOv3, uses EfficientNet-B0 network to replace the backbone network of the original YOLOv3 algorithm, which can extract image features more effectively, increase the size and number of prediction branches and optimized prior boxes, and improve the detection effect of remote sensing small targets. At the same time, DIoU is selected as the loss function to improve the efficiency of target detection box regression and reduce the omission ratio. Experimental results on DOTA remote sensing image dataset show that the mean average precision mAP of the proposed algorithm is 91.01%, which is 11.82% higher than that of the original YOLOv3. Therefore, it has higher detection accuracy

Key words: remote sensing image, object detection, YOLOv3, EfficientNet, multi-scale detection