Learning novel object parts model for object categorization

We present a new method to learn the model based on object parts extraction and grammar which can be applied to classification and recognition. Our approach is invariant to the scale and rotation of the objects. We use Structural Context feature to detect object parts. It is done comparing SC histograms of the model and image. We extract oriented triplets from centers of detected parts. We define grammar for these parts using normalize distances and angles between them. We propose and compare two alternative implementations using different classifiers: Hidden Markov Model with mixture of Gaussian outputs (MHMM) and Adaptive Neuro-Fuzzy Inference system (ANFIS) to learn this grammar and estimate parameters of the model for each object class. The proposed method is computationally efficient and it is invariant to scale and rotation. Experimental results demonstrate the privileged performance of the proposed approach against other methods. © 2010 IEEE.