Location Sensitive Regression Algorithm for Multi-oriented Scene Text Detection with Focal Loss

Scene text detection methods based on direct regression are always efficient and accuracy. However, the previous methods also have obvious shortcomings. The central text pixels always constrained the text box in the process of direct regression, which made the predicted results inaccurate. To solve this problem, we propose a novel Location Sensitive Regression algorithm for multi-oriented scene Text Detection (LSRTD), which assigns different weights to text pixels according to their distance to text box vertexes. At the same time, we apply Focal Loss to pixel-wise text/non-text classification task, the number of false positive cases in results is decreased because of Focal Loss balances the loss weights of easy and hard samples during training. Our method LSRTD is validated on benchmark datasets. The F-measure of ICDAR2015 is 0.84 on single-scale testing.