Abstract: The effect of deep learning depends heavily on the quantity and quality of data,and insufficient data will cause the mo-del to overfit.In practical application research,it is often difficult to obtain a large number of high-quality sample data,especially image data.In response to the above problems,this paper takes the ANIMAL-10 dataset as the research object,and designs a method based on foreground target extraction and using pure colors to replace the original background to achieve data augmentation.Combining traditional data augmentation methods to construct new datasets,four neural network models of AlexNet,Inception,ResNet and VGG-16 are used to analyze the impact of different color backgrounds and different data augmentation methods on the accuracy of model recognition.Experiments shows that different color backgrounds have no significant influence on the accuracy of model recognition.And on this basis,the green background is used for subsequent data augmentation operations,four datasets A,B,C,and D are designed and the above four models are compared.The test results show that the model have a significant impact on the recognition accuracy,while the data set has no significant impact on the recognition accuracy.However,for the AlexNet and Inception-v3 models,the recognition accuracy of the enhanced dataset including the prominent foreground data is increased by 3.78% and 4.55%,respectively,compared with the original image and traditional data augmentation methods.This shows that under small datasets,the data augmentation method that highlights the foreground could make the model easier to notice and learn the key features of the images,so that the performance of the model is better,and the recognition accuracy of the model is improved,which has certain practical value in actual engineering applications.

Key words: Convolutional neural networks, Data augmentation, Deep learning, Highlight the foreground, Image recognition