融合CTPN和CRNN模型的自然场景文本检测与识别方法

A Natural Scene Text Detection and Recognition Method Based on Combining CTPN and CRNN Models

针对自然场景中文本检测与识别存在准确率不高和效率不佳的问题,提出了一种融合场景文本检测CTPN和卷积循环神经网络CRNN模型的自然场景文字检测与识别方法。首先利用CTPN模型对文本行进行预测,再利用卷积神经网络进行特征序列提取和识别并基于Bi-LSTM学习序列特征,最后将文本分数高的窗口连接作为文本区域,从而实现文本检测。实验结果表明,在ICADR-2013数据集中改进模型的准确率可达78%;基于CRNN的文本识别模型在提取特征序列后,经过转录层预测的文本序列结果,在测试集上的准确率可达到86.7%;以上结果表明融合CTPN和CRNN模型的自然场景文字检测与识别方法能够获得更好的检测与识别效果。

A natural scene text detection and recognition method that combines scene text detection CTPN and convolutional recurrent neural network CRNN models is proposed to address the issues of low accuracy and low efficiency in text detection and recognition in natural scenes.Firstly,the CTPN model is used to predict text lines.Then,convolutional neural networks are used to extract and recog⁃nize feature sequences,and sequence features are learned based on Bi-LSTM.Finally,windows with high text scores are connected as text regions to achieve text detection.The experimental results show that the accuracy of the improved model in the ICADR-2013 dataset can reach 78%;the text recogni⁃tion model based on CRNN can achieve an accuracy of 86.7%on the test set after extracting feature se⁃quences and predicting the text sequence results through the transcription layer.The above results indi⁃cate that the fusion of CTPN and CRNN models for natural scene text detection and recognition can achieve better detection and recognition results.