Test Data Augmentation for Image Recognition Software

Image recognition software has been widely used in many vital areas, so it needs to be thoroughly tested with images as test data. However, for some special areas, such as medical treatment, there are only a few sufficient and credible test data. Some test data still depends on the training data, which results in the defect detection ability of the testing is not high. In this paper, we propose a new test data augmentation approach with combing domain knowledge and data mutation. Given an image, our approach extracts the features of the recognition targets in this image based on domain knowledge, then mutates these features to generate new images. In theory, our approach could generate high-quality test data, which helps testing image recognition software adequately, and improving the accuracy of image recognition software.