Homographic Class Template for Logo Localization and Recognition

Abstract

We propose a method for localizing and recognizing brand logos in natural images. The task is extremely challenging, due to the various changes in the appearance of the logos. We construct class templates by matching features between examples of the same class to build homographies. An interconnections graph is developed for each class and the representative points are added to the class model. Finally, each class is depicted by the reunion of the suitable keypoints and descriptors, thus leading to a high precision of the proposed logo recognition system. Results show that we outperform the state of the art systems on the challenging Flickr-32 database.

Logo localization and recognition in natural images using homographic class graphs

“…This is one of the improvements over the previous work (Boia and Florea, 2015), where the class graph was unweighted. This modification leads to a noticeable increase of the recognition accuracy.…”

“…Differing from our previous work (Boia and Florea, 2015), Figure 3. The matching process for two crops of the class "Coca Cola": (a) The matching of the descriptors, (b) the remaining pairs after RANSAC, (c) the two crops overlaid after the homography is applied Table 2.…”

“…This modification leads to a noticeable increase of the recognition accuracy. While many previous works describe classes by a single prototype (Romberg and Lienhart, 2013), (Boia and Florea, 2015), here we use potentially two the second modelling color inverted cases. The system automatically detects the classes which present inverted instances, using a class compactness criteria and generates a second model of the class.…”

“…We use the Djikstra algorithm for computing this path. In our previous work (Boia and Florea, 2015) for the unweighted graph used, the shortest path meant literally using the path containing the fewest nodes between the source image and the target, that introduced the smallest error possible. However, this will not also significate the most complete description of the class.…”

“…The current work is developed from the method previously reported by us (Boia and Florea, 2015) with following major technical differences: (1) while in our older work a single template is build per class (single instance learning), here the class graph is processed and, if necessary, a secondary model is automatically build to model the examples depicted in inverted colors (thus implemented as multiple instance learning -MIL); (2) while in our older work (Boia and Florea, 2015) an undirected binary graph was used, here we switch to a weighted graph that permits better homography composition and automatic identification of the nodes cluster thus of the secondary model. Furthermore, here extend the testing and the analysis of the results.…”