Unsupervised local deep feature for image recognition

Abstract

Unsupervised feature learning is an important problem in computer vision and machine learning. Various unsupervised learning methods, e.g., PCA, KMeans, autoencoder, have been applied for visual recognition tasks. Especially, autoencoder has superior image recognition performance, which is usually used as a global feature extractor. To make better use of autoencoder for image recognition, we propose unsupervised local deep feature (ULDF). We collect patches from training images and use autoencoder to transform vectorized patches into short codes. Then these codes are encoded using Fisher Vector with spatial pyramid. Image recognition with ULDF only requires an efficient linear SVM classifier. Finally, we perform multi-scale voting to achieve more accurate classification performance. In the experiments, we evaluate ULDF on the handwritten digit recognition task and the shape classification task, and show that it achieves highly competitive performance comparing to the state-of-the-art image recognition methods. The results suggest that the proposed local deep feature is superior than traditional global autoencoder.