Unsupervised local deep feature for image recognition

Yang Wang, Xinggang Wang, Wenyu Liu

Abstract: Highlights•ULDF is proposed to make better use of autoencoder for image recognition. It is performed on local patches rather than whole images, which helps to scale the algorithm to realistic-sized images.•Owning to the combination with BoW, it is more robust to local translation.•The short codes transformed from vectorized patches using autoencoder have good low-dimensional property. This property makes it consume less time and less memory while clustering. Moreover, short codes make it possible to use finer level spatial pyramid, such as 2 × 2 and 4 × 4.•It only requires an efficient linear SVM classifier and obtains state-of-the-art classify performance on MNIST.

External IDs:dblp:journals/isci/WangWL16