基于CNN的银行卡数字识别方法

Digital recognition method of bank card based on CNN

在拍摄银行卡时,由于受拍摄角度的不确定性、光照条件的复杂性及卡背景的多样性等众多因素的干扰,使得自然拍摄场景的银行卡数字识别算法存在较大挑战。为此,提出一种基于卷积神经网络(CNN)的银行卡数字识别框架。首先,通过投影矫正、边缘检测和形态学等一系列图像处理算法获取目标数字区域;其次,通过增强的数据集训练一个CNN,使用该网络通过滑窗识别获取上述目标数字区域,输出初始银行卡号序列,生成为一个数字曲线图;最后,提出了滑窗优化算法,该平滑算法输入上述初始的银行卡号曲线图,对其进行优化,继而分割出单个数字并输出最终结果。实验结果表明算法显著提高了银行卡数字识别和分割的准确率,同时针对较复杂的银行卡图像仍然具有较好的鲁棒性。

Due to many interference factors when photographing the bank card,such as the uncertainty of shooting angle,the complexity of lighting conditions and the diversity of bank card background,there are great challenges for the bank card digital recognition algorithm based on natural shooting scene.Therefore,a framework for bank card recognition is proposed based on convolution neural network(CNN).Firstly,the digital region of target bank card is obtained by performing a series of image processing algorithms,such as projection correction,edge detection,and morphology operation.Secondly,a convolution neural network is trained through the augmented dataset to obtain the above target digital area for sliding window recognition.Then the initial bank card number sequence is output to generate a digital graph.Finally,a smoothing optimization algorithm is proposed,which inputs the above initial bank card number graph and optimizes it.Then the digital sequence is divided into individual numbers and the final result is output.The experimental results show that the algorithm significantly improves the accuracy of bank card digital recognition and segmentation.At the same time,it still has good robustness for those bank cards with more complex images.