Weakly-supervised TV logo detection

摘要

In this paper, a TV logo detection system is proposed based on the deep learning architecture for the specific TV logo detection task. Training a robust object detector typically requires a large amount of manually annotated data, which is time-consuming. To reduce the cost, we construct a TV logo detection system in a weakly-supervised framework, which is accomplished by a TV logo localization network based on Region Proposal Network (RPN) and a classification network based on Fast RCNN. Based on observed priors of a typical TV logo in pictures and video frames, data preparation and processing are performed by carrying out keyframe extraction and data augmentation. Since we build the localization network based on RPN, only a few bounding box annotations are employed for training the localization network. Then the well-trained localization network can produce numerous positive and negative proposals. These proposals along with the logo class labels for classification network training are exploited to train the classification network. To generate reasonable anchor boxes, k-means clustering is utilized to infer the scales and aspect ratios. Besides, for efficient training and better generalization ability, hard example mining is also explored. Experimental results demonstrate that the proposed weakly-supervised TV logo detection system achieves superior performances compared to the baseline Faster RCNN approach, with a mAP as about 92% in our newly proposed dataset.