Store classification using Text-Exemplar-Similarity and Hypotheses-Weighted-CNN

Store classification is a challenging task due to the large variation of view, scale, illumination and occlusion. To efficiently distinguish different stores, we introduce two features: Text-Exemplar-Similarity and Hypotheses-Weighted-CNN. For the first feature, the similarity with the discriminative characters is used to represent the text information. For the second feature, we first generate a set of object hypotheses. Then, we introduce two priors: edge boundary and repeatness prior to give a higher weight to the hypotheses enclosing the object. After the generation of two features, a simple and efficient optimization method is used to find the best weight for each feature. Extensive experiments are evaluated to verify the superiority of the proposed method. We built a new 9-class store dataset composed of photos and images from the internet. The experiments show that our method is nearly 10% higher than the state-of-art methods.