This thesis addresses the problem of image-based logotype detection and recognition. A new algorithm for logotype detection in images of cars is proposed. In the first stage, the algorithm localizes all maximally-stable extremal regions as candidates of logotype parts. In the next stage, the regions are combined to create logotype candidates, which are encoded by histograms of gradients. A random forest classifier is then used to verify the candidate regions as being logotypes or not and simultaneously classify them into the type of the logotype. In addition, improvements to basic algorithm are proposed. The improvements include the use of alternative color spaces, geometric normalization and use of the car license plate location to form the logotype position prior. An annotated dataset with photographs of vehicles of twenty different makes was prepared for evaluation of the algorithm. The algorithm was able to correctly localize and recognize over 70% of car logotypes at a very low false positive rate. Despite the fact that we focus on car logotype detection, the algorithm can be easily extended to detection of arbitrary logotypes or objects that do not violate assumptions we impose on the logotype appearance.
|
