In recent years, deep learning has made a breakthrough in image recognition. However, it often requires a large amount of label data as the sample set. In most practical applications, the neural network is prone to over-fitting or weak generalization due to the lack of annotation data. This phenomenon is especially obvious in a small-scale data set. To solve this problem, pine wilt disease data is used as an example to adopt mirroring, flipping, adding noise, rotating, scaling, and other augmentation methods to enhance the amount of the image sets. It can not only increase sample diversity but also make the network more stable for training. In this paper, the effects of different amplification methods and training samples size on the Faster R-CNN and YOLOv3 models are tested, and its results show that scaling has the greatest impact on the two models for the reason that the two models are both sensitive to the size of sample images. The accuracy of Faster R-CNN starts to decline when the number of training sets is expanded to 60% of the new training samples, the accuracy of YOLOv3 starts to decline when the number of training sets is expanded to 75%.