Research on Data Augmentation for Object Detection Based on X- ray Security Inspection Picture

Abstract

The deep learning model has a large number of parameters that need to be adjusted during the training process, so training a model that can perform well in object detection requires a large amount of valid data. And the accuracy and generalization ability of the model are guaranteed by both quantity and diversity of data. However, it’s a huge expense to get pictures and then annotate them, and we can hardly ever get all types of data that exist in reality. Apparently, data augmentation is an effective way to solve this problem. In our paper, the data set we use is composed of x-ray pictures come from subway security inspection. Based on the characteristics of x- ray pictures, we conducted experiments on a variety of augmentation methods, aimed at finding suitable data augmentation methods for this data set. After the experiment, it is concluded that the augmentation method that conforms to the actual form of the data set is more effective. Compared with the monotonous data augmentation, the hybrid way can be more effective, which can improve the accuracy and generalization ability of the model better. Meanwhile the complexity of backbone can also affect data augmentation, more complex backbone can be more sensitive to data augmentation.