RPN with the Attention-based Multi-Scale Method and the Adaptive Non-Maximum Suppression for Billboard Detection

Billboard detection is a special application of general object detection. Although recent object detector algorithms based on deep learning have achieved well performance for general object detection, they have limited success for specific applications. In this paper, a novel method based on Faster R-CNN for billboard detection is proposed. The proposed architecture is called attention-based multi-scale feature fusion region proposal network (AM-RPN). The contribution of this paper is to propose the attention-based multi-scale feature fusion method for billboard detection. Since the billboards and background features from the lower convolution layers are similar, it is difficult to improve detection accuracy using the multi-scale methods which contain the low-level features. The proposed attention-based multi-scale feature fusion method uses attention mechanism to give different focus to the information which constitutes the multi-scale fusion feature. Attention mechanism ensures that the important information is left and the useless information is discarded. Moreover, this paper proposes the adaptive non-maximum suppression (ANMS) algorithm to replace the classical non-maximum suppression algorithm. The adaptive non-maximum suppression algorithm considers the relationship between the classes and allows more low confidence targets to be detected without increasing error detection. Our proposed approach is evaluated on Baidu's public billboard detection data set. Experimental results show that AM-RPN and the class suppression obtains significant improvements over the comparable state-of-the-art detection models for billboard detection.