Multi-oriented text detection from natural scene images based on a CNN and pruning non-adjacent graph edges

Due to the complex backgrounds, size variations, and changes in perspective and orientation in natural scene images, detecting multi-oriented text is a difficult problem that has recently attracted considerable attention from research communities. In this paper, we present a novel method that effectively and robustly detects multi-oriented text in natural scene images. First, the candidate characters are generated by an exhaustive segmentation-based method that can extract characters in arbitrary orientations. Second, a convolutional neural network (CNN) model is employed to filter out the non-character regions; this model is also robust to arbitrary character orientations. Finally, text-line grouping is treated as a problem of pruning non-adjacent graph edges from a graph in which each vertex represents a character candidate region. To evaluate our algorithm, we compare it with other existing algorithms by performing experiments on three public datasets: ICDAR 2013, the Oriented Scene Text Dataset (OSTD) and USTB-SV1K. The results show that the proposed method handles any arbitrary text orientation well, and it achieves promising results on these three public datasets.