基于深度卷积网络的非接触式掌纹识别与验证

Recognition and verification of contactless palmprint based on deep convolutional network

针对非接触式掌纹图像存在手姿态、光照等干扰因素的问题,提出了使用深度卷积网络来提取非接触式掌纹特征的识别方法,对不同网络提取非接触式掌纹特征的性能进行了验证.为了提高实用性,避免非接触式掌纹验证前的ROI提取操作,提出了基于Siamese Network的非接触式掌纹验证方法.选用了ResNet、DenseNet、MobileNetV2和RegNet 4个卷积神经网络模型,在IITD、Tongji和MPD 3个非接触式掌纹数据集上做了非接触式掌纹识别的评估实验,在IITD数据集上进行了训练和验证.MobileNetV2在IITD数据集上的收敛速度最快,RegNet在Tongji、MPD两个数据集上的收敛速度明显快于另外3个网络.RegNet在3个数据集上的识别率均最高,且较传统方法有所提高.实验结果表明,用深度卷积网络提取非接触式掌纹特征的方法有更好的识别结果.基于Siamese Network的非接触式掌纹验证方法对自然场景下的掌纹图像有较好的验证结果,且对光照和手姿态具有一定的鲁棒性.

The unique collection method of contactless palmprints has been widely recognized by the public.However,the contactless acquisition method will cause interference factors such as hand posture and illumination in the contactless palmprint image.To solve this problem,a recognition method using deep convolutional network to extract contactless palmprint features is proposed.Considering that there are relatively few researches on deep learning in contactless palmprint recognition,this paper also verifies the performance of different networks in extracting contactless palmprint features.In addition,in order to improve the practicability and avoid the ROI extraction operation before the contactless palmprint verification,a contactless palmprint verification method based on the Siamese Network is proposed.Four convolutional neural network models,ResNet,DenseNet,MobileNetV2 and RegNet,were selected,and the evaluation experiments of contactless palmprint recognition were done on the three contactless palmprint databases of IITD,Tongji and MPD.The proposed contactless palmprint verification method based on Siamese Network was trained and verified on the IITD database.MobileNetV2 has the fastest convergence speed on the IITD database,and the convergence speed of RegNet on the Tongji and MPD databases is significantly faster than the other three networks.RegNet has the highest recognition rate on the three databases,and it has improved compared with traditional methods.After training on the IITD database,the contactless palmprint verification method based on the Siamese Network can directly verify the palmprint images taken by the smartphone,and the verification results are good.The experimental results show that the method of extracting contactless palmprint features with deep convolutional network has better recognition results.The contactless palmprint verification method based on Siamese Network has good verification results for palmprint images in natural scenes,and has a certain degree of robustness to illumination and hand posture.