Logo detection using weakly supervised saliency map

Box level annotation of a large number of logo images for training purpose of typical deep learning architecture is highly challenging. Thus, a method that can detect the logo with the help of training to remove box-level annotations can be helpful. In this paper, we present a method of logo detection that utilizes weakly supervised learning of Convolutional Neural Network (CNN) to generate a deep saliency map. The saliency map is generated from the back-propagated response of the CNN trained with the classification task. The saliency map produces responses for the regions of logos. GrabCut segmentation method has been applied then to obtain the bounding box corresponding to the logo class predicted by the CNN for a given image. AlexNet, CaffeNet, and VGGNet deep architectures has been fine-tuned for the classification purpose. The framework is further utilized for detection through a back-propagated saliency map. The performance of the proposed methodology has been validated on the FlickrLogos-32 logo benchmark dataset. The proposed method outperforms the state-of-the-art baseline fully supervised methods with mean average precision (mAP) of 75.83%.