An Xception Based Convolutional Neural Network for Scene Image Classification with Transfer Learning
Xizhi Wu, Rongzhe Liu, Hanqing Yang, Zizhao Chen
Abstract
Over the past decade, image classification, which can provide assistance to address complex tasks such as planetary exploration and unmanned driving, has become a hot topic. As a subproblem of image classification, scene image classification has received increasing attention. Based on previous studies, the Xception model achieved superior performance on image classification tasks in comparison with the original Inception model. The Xception model is advantageous at processing image classification, yet it has not been used for scene image classification. To tackle this issue, this paper proposed an Xception based transfer learning, and analyzed the model performance by comparing it with the Inception-V3 model. We found that the Xception based transfer learning significantly outperforms other methods such as Inception-V3, which is nicely demonstrated by the experimental results on the Intel Image Classification Challenge dataset. Furthermore, the Xception has shown greater robustness and ability in generalization with less overfitting problems.