Research On Near-infrared Iris Detection Algorithms

Iris detection is the base and key of iris recognition technology.Although there are many good iris detection methods,their stability is still vulnerable to complex background,occlusion,uneven illumination and other factors under unconstrained scenes.This paper focuses on researching the near-infrared iris detection problem under unconstrained scenes,so we propose two iris detection algorithms.One is based on LRND and DTS,the other is based on Refined Mask RCNN separately.The details are as follows:The iris detection algorithm based on LRND and DTS is proposed to mainly solve the iris regional detection problem.First,we propose a new image feature named Local Region Based Normalized Difference(LRND)feature and its corresponding iris regional representation method.Then,Dual Threshold Stump(DTS)is used as the weak classifier and will be combined with Gentle AdaBoost algorithm and Soft Cascade structure to generate strong classifier.To solve multi-scale iris detection problem,we construct a feature pyramid.For overlapping problem among multiple adjacent boxes,Non-Maximum Suppression algorithm is adopted during post processing.Experimental results show that this method achieves 99.45%average recall and 99.35%average precision on CASIA Distance and MIR-Train binocular iris databases.Each image takes about 0.04s.These results prove this method has high performance on regional detection.The iris detection algorithm based on Refined Mask RCNN corresponds to a multitask integration model,which can solve the problem of regional and pixel level detection at the same time.The model optimizes the original Mask RCNN framework.It resolves the region inconsistency problem between predicted mask and detection bounding box by adjusting the input of instance segmentation network and model learning method.In addition,we present a series of model enhancement schemes as follows:(1)We propose a network structure optimization based on stage residual block for feature backbone network to expand the feature representation space.(2)We propose a Chain-cascade Atrous Spatial Pyramid Pooling(CASPP)structure.It successfully achieves pixel level dense prediction and class based semantic segmentation from a macro perspective.(3)We also use large mini-batch learning and online data augmentation to enhance model training and improve the adaptability of models to complex scenes.Experiments show that the method obtains 99.61%regional recall,100%precision on CASIA binocular database and takes about 0.260s per image during testing.The method also obtains 99.83%regional detection recall,100%precision,96.96%pixel level detection mIOU on CASIA Thousands monocular database and takes about 0.178s per image during testing.These results prove this method has strong multi-function detection performance.In conclusion,this paper starts from the route of traditional machine learning and deep learning separately,proposes two corresponding near-infrared iris detection algorithms.Both of them achieve successful iris detection with high efficiency and high precision.