Correction On Fingerprint Orientation Field And The Evaluation Of Image Quality Based On Deep Learning

With the rapid development of information technology,fingerprint identification,as a representative biometric identification technology,has been widely used in various security fields.In practice,the processing effect of fingerprint image will directly affect the accuracy of fingerprint identification.In fingerprint image processing,many operations are carried out based on fingerprint direction field.Therefore,the accuracy of direction field calculation is especially important for the accuracy of subsequent processing.In practice,the collected fingerprint image will be interfered by all kinds of noises,resulting in the wrong calculation of the direction field,so it is necessary to correct the disordered part of the direction field.Theoretically,this problem can be classified as image denoising.Traditional image denoising methods denoise specific types of noise,while fingerprint image noise is uncertain and random,It is difficult to solve the problem of directional field correction effectively by using deterministic algorithm.So far,deep learning has demonstrated a strong ability to learn in all walks of life and to deal with complex problems.Therefore,this paper tries to use deep learning to solve the problem of direction field correction.First,make a training set.Using the existing methods to calculate the fingerprint direction field as the original experimental data,0 to 7 were used to represent the 8 directions of the fingerprint,and selecting a part of relatively clean data set from the original data set,without disordered direction block in the image center as the standard data set,Then,this part of data is artificially and randomly disturbed to make the noise data set;Secondly,construct the deep convolutional neural network.The noise data and corresponding standard data are fed into the network in pairs for supervised training,And using the idea of residual structure,the network can learn the distribution of noise automatically,The direction field data after denoising is obtained by subtracting the generated noise data from the original direction field data with noise;Finally,the data representation of fingerprint direction field is analyzed and improved by experiments.After the network training test,it is found that there are still some disordered direction block data in the results obtained by using 0 to 7 to represent the direction field data.In order to consistent with the periodicity of fingerprint direction field,using two-dimensional data(cos?-???/4,sin4^-??/4)to represent the direction of the field data.And through the experiment,the direction block which can correct the disorder well.In the process of macro features and details of the fingerprint image extracting,characteristics need to give a degree of confidence to authenticate subsequent identification,which requires an image quality evaluation system.The quality of the fingerprint direction field basically reflects the quality of the corresponding fingerprint image usually.Thus the evaluation of fingerprint image quality is transformed into the evaluation of fingerprint direction field quality.The basic architecture of the evaluation network we built is the convolutional neural network,in order to strengthen the noise information,every time will put the field data which need to be evaluated and the corresponding field data which have been denoised after denoising network in pairs into the network,as the initial input of network,and then add the noise data and the data need to be evaluated as the new input of network.According to the different disturbance degree to different grade quality of fingerprint direction field,there are 0.2,0.4,0.6,0.8 as labels,training the evaluation network with supervision.Putting the original directional field data into the evaluation network and getting its score.In practical application,even if a large network can be constructed to realize denoising of the direction field data with poor quality,the corrected information is not be obtained in the image processing and feature extraction of the subsequent direction field.Therefore,the data with a relatively low score is regarded as"invalid data",and the data with a relatively high score is corrected effectively.The evaluation network and denoising network are interacted,the availability of de-noising data in practical application can be guaranteed.When the denoising capability of the denoising network is further improved,the score of the corresponding"invalid data"can be reduced appropriately.This method is of great significance to the subsequent fingerprint image processing.