Dorsal Hand Vein Recognition Based On Concept Learning And Deep Learning

Nowadays,the technology of hand vein recognition has made it more suitable for people's use psychology because of its unique characteristics.But limited by the external factors such as region,posture,and equipment,there are two extreme cases of "large sample" and "small sample" in the number of images of the dorsal vein image,and the corresponding identification methods and strategies are different.Therefore,this paper puts forward a "concept learning" strategy of "small sample"hand back vein identification,and a "deep learning" strategy of "large sample" hand back vein identification,and the two cases are tested respectively.The main work and innovation of this paper are as follows:(1)In view of the low efficiency of the traditional BOW model(Bag of word,BOW)in the clustering process and the poor performance in the class,the "concept learning" method based on the K-means++ multi cluster model is proposed for the recognition of the dorsal vein of the hand.Mainly including the use of the Dense SIFT algorithm to generate geometric primitive description of the vein lines,and then using the improved K-means++ multi clustering algorithm to cluster each cluster center together and use the mutual information algorithm to remove the redundancy to build the geometric component set,so as to improve the efficiency and highlight the performance in the class.Finally The spatial Pyramid model is used to complete the recognition by mapping the primitive map to the histogram feature of the component set.(2)In view of the differences in the rotation and brightness of the large sample hand vein images collected under the weak constraints of cross equipment,cross region and cross time,a "depth learning" method of Center Loss combined with the loss function of Softmax Loss is used in the AlexNet model.A reasonable image training set is constructed to make the variety of feature changes in the training set,and then the Center Loss layer is added to improve the cohesiveness of a variety of change features to solve the problem of image differences in cross equipment,cross region and cross time.(3)Verify the two methods based on "concept learning" and "deep learning"based on the actual collection of the hand vein library.The experimental results show that the improved K-means++ multi clustering model is up to 32%and the recognition rate is up to 85.714%compared with the traditional BOW model.Compared with the original model,the recognition rate of the AlexNet model with Center Loss layer has been further improved to 92.6%.In this paper,the problem of back vein identity identification is studied in two extreme cases.It is verified by experiments that the two methods have better recognition performance.