Research On Non-contact Palmprint Recognition Algorithm

Technology of biometric identification is playing a more and more important role in activities of production and people's lives.It is contributed more to recognition of identity.Palmprint recognition has caught a great attention due to its stability,uniqueness,accuracy and reliability.Although current palmprint recognition methods have successfully applied in many fields,most of them use touch devices.Images collected by this kind of devices are influenced less by illumination,scales or other factors,but they should be improved on user friendliness.Focusing on the mentioned problems,in this paper,we use touchless palmprint images which is affected seriously by non-uniform of illumination,varying scales,rotation and deformation.Then the recognition accuracy is surely influenced,the cost of the whole recognition system is large because of storage of palmprint feature database.So,we proposed a feature extraction method of Mixture of Sobel-Gabor(MSG)to exactly describe the palmprint information on scale and orientation through the combination with blanket dimension(BD)algorithm.We also proposed a method based on classification thought,the recognition rate is improved as well through support vector machine(SVM)and scale invariant feature transform(SIFT),to the cost of system,we have avoided the storage of features.Finally,we proposed a batch-wise(BS)training SVM method,several SVM classifiers are trained according to images sets under different scales.When a testing sample is input to the recognition system,there will be several results,then the ultimate result is decided by chi-square distance and weighted mean.The main works include following aspects:(1)Conduct Sobel transform to all samples,then eliminate samples those which may brought about fake features via two kinds of Gabor transform and Minkowski distance,extract BD features from obtained ultimate Gabor feature maps and use correlation computation to make a classification;(2)We proposed a classification based palmprint recognition method which is able to avoid the cost caused by storage of palmprint features.Firstly,extracting SIFT features for all training samples,then put into the SVM classifier,the obtained classification results is determined by chi-square distance;(3)In this part,we proposed a batch-wise training method.Firstly,conduct a scale transform for all training samples,then we got some training sets in different scales,training these sets respectively and several SVM classifiers is generated.A testing sample will be put into all classifiers obtained,and ultimate result is determined by chi-square distance and weighted mean.