Unsupervised Image Classification System for Mobile Devices Based on Incremental Isomap

Unsupervised Machine Learning methods could classify massive images with considerable accuracy, without sensitivity to the selection of labeled data sets compared with supervised learning methods. However, once facing a huge number of images, feature extraction and dimensionality reduction during unsupervised process tend to be more complex and therefore hard to run on mobile devices in real time. In this paper, we propose an unsupervised method based on Incremental Isometric Mapping algorithm integrated with BagNet models for image classification on mobile devices. Then we establish a new image classification system and utilize cloud computing for fast computation so that the system could realize real-time image classification on mobile devices. Finally, we deploy the proposed system on commercial mobile phones and achieve an classification accuracy of 77.9% on photos of users. With the help of the cloud, time cost of the system has declined from 14.55s per image to 1.08s per image.