A New Method of Image Detection for Small Datasets under the Framework of YOLO Network

For the image detection problems under smallscale datasets, the detection rate of deep learning method is usually very low. This paper presents a new image detection method based on transfer learning and sample enhancement under the framework of YOLO network. This method takes full advantages of the real-time feature of YOLO, as well as the enhancement of the generalization ability brought by transfer learning and sample enhancements. The detection rate of 87.4% of the 6 targets under the small-scale datasets was achieved, and this method is more than 6 times faster than the Faster R-CNN at the same detection rate. The measured data verified the method. Furthermore, this paper quantitatively analyzes the relationship between sample scale and detection performance under smallscale datasets.