Semi-supervised object recognition based on Connected Image Transformations

We present a novel semi-supervised classifier model based on paths between unlabeled and labeled data through a sequence of local pattern transformations. A reliable measure of path-length is proposed that combines a local dissimilarity measure between consecutive patters along a path with a global, connectivity-based metric. We apply this model to problems of object recognition, for which we propose a practical classification algorithm based on sequences of “Connected Image Transformations” (CIT). Experimental results on four popular image benchmarks demonstrate how the proposed CIT classifier outperforms state-of-the-art semi-supervised techniques. The results are particularly significant when only a very small number of labeled patterns is available: the proposed algorithm obtains a generalization error of 4.57% on the MNIST data set trained on 2000 randomly chosen patterns with only 10 labeled patterns per digit class.