Object Detection Using Signature Library

This paper describes a novel method of detecting an object using signature library which is very fast as compared to traditional methods. The traditional method deals with template matching based object detection. It involves sliding window based approach, which is very costly and time consuming. The template matching based object detection method searches for the object in all possible image sub-windows. Based on resolution of the image huge numbers of sub-windows are extracted, which are used for training the classifier. Our method tries to overcome the problem of extracting every possible sub-window from the images and training the classifier with the extracted sub-windows by constructing signature library from selected categories of images. We are using the popular Bag-Of-Words (BoW) paradigm for extracting features from images. For Signature library generation, the BoW feature vectors are extracted from a representative set of images, which are used to learn signature using unsupervised clustering method. This is carried out for each classic specific images. The collection of cluster centres for a set of categories constitutes our signature library. To determine whether the object is present/absent, we first compute the signature for given query image using three level spatial pyramid representation, an extension of bag-of-features representation. The query image feature vector for each spatial resolution is matched with signatures of each class using similarity measure to determine the presence/absence of set of class available in signature library. We will get votes for each class. The votes are sorted in descending order. Finally, the query image is labelled with top matching class, indicating presence of that class. For a given query image, we generate the table which indicate the votes for each class. We have also used SVM for determining the presence/absence of the object. The experimental results are promising with good accuracy indicating the efficacy of our approach. Our method is independent of the resolution of the image, hence it is computationally fast Our visual object detection method can be exploited to the maximum extent if it is deployed in distributed enterprise system.