Object Recognition with Naeive Bayes-NN via Prototype Generation

摘要

Naieve Bayes nearest neighbors (NBNN) is a variant of the classic KNN classifier that has proved to be very effective for object recognition and image classification tasks. Under NBNN an unseen image is classified by looking at the distance between the sets of visual descriptors of test and training images. Although NBNN is a very competitive pattern classification approach, it presents a major drawback: it requires of large storage and computational resources. NBNN's requirements are even larger than those of the standard KNN because sets of raw descriptors must be stored and compared, therefore, efficiency improvements for NBNN are necessary. Prototype generation (PG) methods have proved to be helpful for reducing the storage and computational requirements of standard KNN. PG methods learn a reduced subset of prototypical instances to be used by KNN for classification. In this contribution we study the suitability of PG methods for enhancing the capabilities of NBNN. Throughout an extensive comparative study we show that PG methods can reduce dramatically the number of descriptors required by NBNN without significantly affecting its discriminative performance. In fact, we show that PG methods can improve the classification performance of NBNN by using much less visual descriptors. We compare the performance of NBNN to other state-of-the-art object recognition approaches and show the combination of PG and NBNN outperforms alternative techniques.