为有效区分手写体与印刷体,提出了一种基于卷积神经网络隐层帧特征的分类方法。基于卷积神经网络,提取隐层帧特征,利用高斯混合模型结合隐马尔可夫模型的方法对该特征进行建模,再通过维特比解码算法判定每帧特征的类别。基于帧特征的识别结果,结合文本行图像信息对识别结果进行后处理,确定最终的手写体和印刷体的区域。在签名文书类文本行图像上,相比基线,所提方法对手写体与印刷体分类的识别率提升10.8%和27.57%。在自然场景、表格和带噪文档行验证了其有效性。

To effectively distinguish the handwritten and printed texts, a discrimination method is proposed based on the hidden layer frame features of a convolutional neural network. The hidden layer frame feature is extracted by the convolutional neural network. The Gaussian mixture model is first combined with the hidden Markov model to model the features, and then the Viterbi decoding algorithm is used to determine the category of each frame feature. Based on the recognition results of the frame features, the recognition results are post-processed in combination with the image information. The final handwritten and printed text areas are determined. For the signature document line images, relative to the baseline, the discrimination accuracy of handwritten and printed texts by the proposed method increases by 10.8% and 27.57%, respectively. The effectiveness of the proposed method is verified with the natural scenes, tables and noisy documents.