Scene text detection based on multi-layer fusion and dual-modal perception

• 1、School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi 214122, China

Abstract：Nowadays, convolutional neural network is wildly used in scene text detection, which has greatly improved the results of the scene text detection. However, the dispersion of text distribution and the difference of text scale still bring challenges to the text detection, and pixels in text regions may be mistaken as background in complex scenes, leading to the degradation of detection performance. In order to deal with the problem of large scaling gap between texts, this paper presented a multi-layer weight fusion module. In this module, the deepest feature of backbone is used to generate features and weights with different receptive fields, which makes the output of the module has a comprehensive receptive field to capture text instances with various scales. Besides, this paper put forward a dual-modal perception module to capture the information of the text regions from local and global perspectives of the feature map, which effectively alleviates the problem of the dispersion of text distribution. To solve the problem of background misjudgment, foreground and background enhancement branch is used to strength the supervision of text regions and suppress background noises. Extensive experiments are conducted on three datasets, and the results prove the effectiveness of the proposed method in this paper.

Keywords： deep learning scene text detection neural network complex scenes receptive field background misjudgement