基于WT-BTC特征和SVM组合分类的场景文本检测算法

针对自然场景文本检测在复杂背景下虚警高的问题,提出利用小波变换(wavelet transform,WT)和方块编码算法(block truncation coding,BTC)相结合的方式(WT-BTC)表征文本纹理,并结合支持向量机(support vector machine,SVM)完成对候选文本区域的分类确认。算法首先利用边缘检测和启发式规则快速确定候选文本区域;然后对候选文本区域进行小波分解和BTC编码,提取水平、垂直、对角方向的WT-BTC纹理特征;使用三个SVM分类器分别对不同方向纹理特征学习训练,组合SVM模型实现候选文本区域的二次检测,确认文本区域。实验结果表明算法提高了文本区域检测鲁棒性,在复杂背景条件下对场景文本有较好的检测效果。

Scene Text Detection Method Based on WT-BTC Feature and SVM Combinational Classification

To reduce false alarm rate of the natural scene text detection in complex background, a text texture representation method based on the combination of Wavelet Transform(WT) and Block Truncation Coding(BTC) is proposed, which is combined with the support vector machine(SVM) to classify the candidate text regions. The proposed method firstly uses edge detection and heuristic rules to determine the candidate text regions; Then wavelet decomposition and BTC coding for candidates text regions are used to extract WT-BTC texture feature in horizontal, vertical and diagonal direction, and three SVM classifier are applied for training. Finally, the generated SVM models are combined for detecting candidate text regions secondly to confirm the text regions. The experimental results show that the algorithm improves the robustness of the detection for scene text, and has better detection effect under the condition of complex background.