Impact of a codebook filtering step on a galois lattice structure for graphics recognition

In this paper, we propose an evaluation of the impact of a codebook filtering step on the recognition rate of a Galois lattice classifier. Unlike the usual approach which only considers a whole visual dictionary and is likely to over-fitting, we boost the Galois lattice using a filtered dictionary by assigning a probability of appearance to each visual word in a symbol model. The retrieval performance and behavior of the method have been compared to state-of-the art and proved that is suitable to the recognition process. Experimental results show that the Galois Lattice classifier combined with a filtered codebook outperforms classic classifiers. Interestingly, due to the high selection of features from the dictionary, the accuracy improvement is obtained with a considerable computational cost reduction.