Research On High Resolution Fingerprint Recognition Algorithm Based On Sweat Pore Feature

In today's information age,more and more biometrics are used in social affairs.Fingerprint is one of the most commonly used biometrics which is highly discriminating and unique.Although it has only a small piece of skin,but can provide quite large information for identification.Compared with other biological features such as faces,fingerprints are permanent and will not change along with the growth of the age.Therefore,fingerprint recognition has become the most commonly used biometric authentication technology for human identity verification in our daily life.Can also provide convenience for social affairs such as judicial investigation,identity verification and device unlocking.In recent years,with the increasing demand for fingerprint acquisition equipment and fingerprint matching accuracy,the effect of traditional low-level fingerprint features is difficult to achieve considerable results.In order to achieve higher accuracy,we used the most subtle feature named sweat pore in the fingerprint for our research studies.However,there are still some problems and challenges in pore extraction and matching.This paper is aim to solve the existing problems in the existing pore matching method in order to achieve better results.First,in the method of extracting sweat pores,in order to better adapt to the shape and size of sweat pores,this paper proposes a method combining deep learning and traditional image processing technology to detect pores.The method first performs a pre-detection on the place where there may be pores in the fingerprint image,and then uses the well-trained Convolutional Neural Networks(CNNs)model to judge these positions.Finally,the results are refined by a series of morphological processing.After these operations,we believe that the method can provide a high recall rate while maintaining a low false positive rate.The experimental results also show that our method is more robust to fingerprint images with different gray levels and contrasts.Then,this paper proposes a “distinctive pore” concept which refer to sweat pores are located around the singular point and minutiae in the fingerprint image.Compared with the other sweat holes,we believe that these pores have more special characteristic.Instead of using all the pores,we only use “distinctive pore” to match fingerprint.As a result,the matching time is greatly shortened because the quantity of sweat holes is greatly reduced.At the same time,erroneous reception caused by overlapping of partial regions is also avoided due to the removal of interference from irrelevant pores.The experimental results show that using our “distinctive pore” can greatly improve the matching accuracy.Meanwhile,can also reduce the matching time and calculation amount.Finally,we use a coarse-to-fine matching strategy on the matching sweat pore method.Briefly,CNNs are used to generate a pore representing namely DeepPoreID and used for subsequent matching steps.In generating the pore representation,we first use a sufficient number of different pore patches to train network as classification task.Then use the value of fully connected layer as our pore representation vector.Same to the classical DeepId series method in the face verification domain,we purpose to generate discriminative vectors that produce more similar vectors for similar images,meaning the spatial distance of the vectors is closer and the Euclidean distance is smaller.Because of CNNs' large learning ability,it can generate more robust and effective representation vectors than traditional methods with enough training data.Our experimental results fully demonstrate that our DeepPoreID is robust to grayscale,contrast,rotation and distortion.In this way,our method can be well adapted to the genuine pairs that seems unlikely due to local deformation and the imposter pairs that looks very similar due to the same local ridge flow.Therefore,we can improve the matching accuracy finally.Especially for the fingerprint image with small overlap area,there is an obvious improvement on matching accuracy.