Real-time localization of multi-oriented text in natural scene images using a linear spatial filter

摘要

This paper proposes a multi-oriented text localization method in natural images suitable for real-time processing of high-definition video on portable and mobile devices. Our method is based on the connected components (CC) approach: first, CC are isolated by convolving a multi-scale pyramid with a specifically designed linear spatial filter followed by hysteresis thresholding. Next, non-textual CC are pruned employing a local classifier consisting of a cascade of multilayer perceptron (MLP) fed with increasingly extended feature vectors. The stroke width feature is estimated in linear time complexity by computing the maximal inscribed squares in the CC. Candidate CC and their neighbors are then checked using a more context aware neural network classifier that takes into account the target CC and their vicinity. Finally, text sequences are extracted in all pyramid levels and fused using dynamic programming. The main contribution of the work presented here is execution speed: the CPU-only parallel implementation of the proposed method is capable of processing 1080p HD video at nearly 30 frames per second on a standard laptop. Furthermore, when benchmarked on the ICDAR 2013 Robust Reading and on the ICDAR 2015 Incidental Scene Text data sets, our system performs more than twice faster than the state-of-the-art, while still delivering competitive results in terms of precision and recall.