Research On Text Detection Methods In Natural Scene Images

Text embedded in natural scene images generally provides rich semantic infor-mation,which is of great use for understanding the scene and the enclosed objects.However,to detect text in natural scene images is a challenging task due to the vari-ety of text appearance and complicated scene contexts,and remains a hot topic in the computer vision research area.In this paper,we proposed two novel methods for detecting text in natural scene images:1)text detection by hierarchical localization and growing of textual compo-nents and 2)text detection based on Windowed Maximally Stable Extremal Regions.In the first method,we firstly integrate Maximally Stable Extremal Region algorith-m(MSER)into Stroke Width Transform algorithm(SWT)to extract seed text pixels.Next,we exploit a hierarchical localization and growing strategy on both intra and in-ter character level to extract the characters as completely as possible.We then exploit a Convolutional Neural Network(CNN)to filter out the non-text candidate,and pro-pose a text-line extraction method based on the Random Walk algorithm to aggregate characters into corresponding words.In the second method,we firstly generate a set of MSER smoothness images based on the Maximally Stable Extremal Region algorithm.Then we exploit the sliding window strategy to extract the candidate character regions,named Windowed Maximally Stable Extremal Regions(WMSERs).Next,we filter out three kind of non-text WMSERs:Partial WMSER,Non-Symmetrical WMSER and Noisy WMSER.Further,we extract a set of textual features depicting the relationship between a WMSER and its belonging MSER and train a Random Forest classifier,cascaded by a convolutional neural network(CNN)to filter out non-text.Finally,we aggregate the characters into words using a partition and pruning algorithm.We investigated the effectiveness of the two proposed text detection methods on some public scene text datasets with standard evaluation protocol.The experiment re-sults show that,the proposed text detection method by hierarchical localization and growing of textual component achieves the state-of-the-art results on every dataset(es-pecially the significantly improved recall rate)due to the growing strategy.Meanwhile,the proposed text detection method based on Windowed Maximally Stable Extremal Region is capable of robustly handling characters composed of multiple discrete com-ponents such as Chinese characters.Both methods have the potential to be further improved for higher performance in the future work.