A coarse-to-fine logo recognition method in video streams

Visual logo recognition is significant for many applications, such as enterprise identification, entertainment advertising, vehicle recognition, road sign reading, trademark protection, and much more. In this paper, we propose a coarse-to-fine framework to recognize visual logos from video streams. To reduce the instability of the initial template selection problem, we introduce the “iconic template” selection strategy to select effective template set for visual logos. At the coarse stage, we adopt DOT(Dominant Orientation Templates) matching with a low threshold to find logo candidates. At the fine stage, we transform the multiple template matching problem into a pairwise binary classification problem. The candidates collected from the template matching process combined with the target template are send to a pairwise binary classifier to predict whether the candidate and the template belong to the same logo or not. The pairwise binary classifier is trained in an offline manner and with an unsupervised training data collection strategy. The proposed method can flexibly adapt to different template matching approaches and various matching thresholds. The false-alarm rate is greatly reduced through the second stage. Experimental results show the feasibility and effectiveness of the proposed approach.