A learning-based logo recognition algorithm using SIFT and efficient correspondence matching

In this paper, a novel method for recognizing logos in natural images is presented. The difficulties lie in: (1) the logos in different images are highly variated in shape and location; (2) images will usually have several different kinds of logos; (3) the logo will be occluded by other objects, which traditional methods usually fail. Therefore, a learning-based logo recognition method is proposed to detect and classify the logos in natural image. First, the SIFT matching solution is applied in a set of training data to reliably detect the interest region in images and extract the discriminate features, which is used as the signature of each logo. Second, the approximate nearest neighbor searching strategy is build up by formulating the data into tree-based structure, for the purpose of efficient matching. Finally, to recognize the logos in test image, the corresponding SIFT features will be computed in the interest regions of test image and matched in the database which is achieved in the training stage. Promising results have been obtained in robustly classifying thousands of logos in the images captured by mobile phones, which the recognition accuracy is 95%.