Visual recognition based on randomized visual dictionaries

In visual recognition, Bag-of-Visual-Word(BoVW) method has been widely used due to its excellent categorization performance. However, the conventional BoVW method has problems of visual word synonymy and ambiguity, meanwhile its time efficiency decreases along with visual data scaling up. Thus, this paper proposes a visual recognition method based on randomized visual dictionaries. Firstly, we introduce E²LSH(Exact Euclidean Locality Sensitive Hashing) to cluster local feature points of training video key frames and construct a group of scalable randomized visual dictionaries. Secondly, based upon these randomized visual dictionaries, we train a group of SVM classifiers for each visual concept. Finally, a voting strategy is utilized to integrate opinions of SVM classifiers, thus accomplishing visual recognition. Experimental results show that compared to traditional BoVW method, our method achieves higher visual recognition accuracy, meanwhile guaranteeing acceptable time efficiency.