Automatic flag recognition using texture based color analysis and gradient features

To add to the growing corpus of handy computer vision applications on smart-phones and Tablet PCs, this paper presents our work on real-time Flag Recognition. The novelty of our attempt at recognizing country flags lies in a three-fold contribution - a 56805 flag images database, 38532 for training and 18273 for testing; a generic recognition approach suited not only to flags but also other insignia and object recognition tasks in which the major discriminative information lies in the relative spatial distribution of colors; and practical usability of the approach in a smart-phone application, being trained and tested on a diverse set of camera captured flag images and having real-time performance. With its large number of classes, little or no shape based difference, high inter-class color similarity, and much intra-class color variation, the 224 country-flags database proves to be very challenging. Additionally, the real-time database incorporates considerable variation in texture, scale, illumination and viewpoint. Our work introduces an improved MSD approach incorporating new and revised criteria for HSV based color binning, applies it in a `by-parts' manner and reinforces it with gradient analysis to achieve an accuracy of 99.2% on the training set and an accuracy of 76.4% on the test set. For practical purposes, top-5 and top-10 results accuracies have also been compiled over the test data, reaching 92.46% and 95.56% mark respectively.