Image Classification: A Random Semi-Supervised Sampling Approach

An image classification method is presented based on random semi-supervised sampling (RSSS). RSSS is an iterative algorithm in which the following two steps alternate till convergence:(1)random semi-supervised sampling; (2)semi-supervised spectral clustering for sample labeling and SVM for model training. RSSS uses local spatial histogram as the image feature and it can combine the image spatial relations with statistical information together. The experiments show that the proposed method can use unlabeled images to improve image classification performance and is not sensitive to image geometrical transform.