Learning Robust Scene Classification Model with Data Augmentation Based on Xception

Abstract

Scene classification technology based on computer vision has been widely applied in many fields. However, with the increasing complexity of images, many computer vision classification models are difficult to meet requirements of current scene classification tasks, as they not only require considering the object, background, spatial layout and other information, but also many relationships in the image. Based on the analysis of existing scene classification algorithms and Xception model, an approach that adds optimization from two aspects of data set processing is proposed to solve complicated scene classification tasks. Combined with the image enhancement technology, the serialized image enhancement method is used to expand the dataset and enhance the image features, and takes advantage of the Xception model to extract the image features to obtain the scene classification model with high robustness. The experimental results showed that Xception model was able to deal with scene classification efficiently by making up for the shortcomings of traditional Convolutional Neural Networks (CNN) models in feature extraction and generalization ability.