用卷积神经网络分类最大稳定极值区域实现汉字区域定位

获取对应笔画级连通区的最大稳定极值区域,实施形态学闭操作融合相距较近的最大稳定极值区域,融合后最大稳定极值区域对应的单个汉字区域;利用灰度共生矩阵描述最大稳定极值矩形区域的纹理信息,将其作为卷积神经网络的输入,卷积神经网络对最大稳定极值区域进行分类,过滤非汉字部分;利用最大稳定极值区域颜色直方图的Bhattacharyya距离等特征对最大稳定极值区域进行聚类,同一类最大稳定极值区域组合得到汉字文本候选区域;再次利用卷积神经网络对候选文本区域进行分类,过滤非文本部分,剩余的就是定位到的汉字文本区域。实验结果表明,该算法对于汉字区域定位具有良好的效果。

Scene Chinese text localization by convolutional neural network classifying maximum stable extremal regions

In order to recognize Chinese text in natural scene image, an improved algorithm for locating Chinese text based on convolutional neural network classifying maximum stable extremal regions is proposed in this paper. Firstly, it extracts the maximum stable extremal regions(MSERs) which corresponds to Chinese strokes. The morphological close operation is used to connect the nearby MSERs. The fused MSER corresponds to Chinese character. Gray level co-occurrence matric is used to describe the textural characteristics of fused MSER rectangle. They are the input of Convolutional Neural Network (CNN). The MSER rectangles are classified by CNN in order to filter none Chinese character rectangle. Then, Chinese text candidates are constructed by clustering MSER rectangles based on the features such as the color histogram Bhattacharyya distance of MSER rectangles. CNN is used to classify Chinese text candidates to filter none Chinese text clusters. Finally, the rectangle of remain clusters are the Chinese text regions of natural scene image. Experiment shows that the proposed algorithm is excellent in localizing the Chinese text in natural Scene images.