An Orientation-correction Detection Method for Scene Text Based on SPP-CNN
Jin Dai, Liang Guo, Zu Wang, Shuguang Liu · 2019
Scene text detection plays an important role in computer vision and pattern recognition field in recent years due to extract the accurate and rich text information. At present, component-based methods have become the trend, and there are still some challenges because of illumination, blur and difficult background. In this paper, an orientation-correction detection method for scene text based on Spatial Pyramid Pooling Convolutional Neural Networks, SPP-CNN , is proposed. Firstly, an enhanced multi-channel MSER model, which is constructed from R, G, B, H, S, V and grey channels of manually blurred and sharpened images, is built. Then, the manually-designed features are embedded in SPP-CNN as the effective text feature detector and classifier. Finally, a two-layer text grouping algorithm is achieved which can handle slightly-slanted text. Experiments on ICDAR 2011 show the f-measure has improved to 84%, the precision and f-measure of ICDAR 2013 have reached 87% and 85% respectively.