A robust hierarchical detection method for scene text based on convolutional neural networks

Abstract

Detecting the text in natural scene images is often challenging due to the complexity and variety of text's appearance and its interaction with the scene context. In this paper, we present a novel hierarchical text detection method exploiting textual characteristics at both character and text line scales for improved accuracy. First, seed candidate characters are detected with discriminative deep convolutional features learned within the maximally stable extremal regions extracted from the image, and are further grown to localize other degraded candidate characters. Next, as a finer filtering of text in the richer text line context, the random forest classifier is exploited on statistical features of text line characterizing the geometrical and conformability properties of constituent character components, to predict the text and non-text label. The effectiveness of the proposed method is demonstrated by the state-of-the-art results achieved on the public datasets.