为优化手写签名识别算法性能,提出了一种局部二值模式(LBP)和深度学习相结合的手写 签名识别算法。针对签名图像进行预处理、维纳滤波去除噪声;将预处理签名图像分为3×4子 块,LBP应用于分块后的每个子图像,并将每个子块的纹理直方图特征连接起来,形成全局直方 图特征;将得到的特征向量作为深度信念网络(DBN)的输入,逐层训练网络,并在顶层形成分类 面,对签名图片进行识别。基于GPDS、MCYT及原创数据库进行实验,识别率误差分别为5.85%、 9.3%、1.17%,有效提高了手写签名的识别精度,符合实际应用的要求。

In order to improve the performance of handwritten signature verification algorithm, a handwritten signature verification algorithm based on local binary pattern (LBP) feature and deep learning is presented. Aiming at signature image, preprocessing and Wiener filtering are used to get rid of noise. The preprocessed signature image is divided into 3×4 blocks and LBP is used to each sub-block. The texture histogram characteristics of each sub-block are connected to form a global histogram characteristics. The obtained feature vectors are used as inputs of deep belief network (DBN), grid is trained layer by layer, and the classification plane is formed at the top to recognize the signature image. Experiments are conducted based on GPDS, MCYT and the original database, and the recognition rate errors are 5.85%, 9.3% and 1.17%, respectivly. The handwritten signature recognition accuracy is effectively improved, which meets the requirements of practical application.