A Study Of Internal Knuckle Pattern Recognition Method With Fusion Feature Encoding

The demand for contactless hand recognition in personal identification is increasing day by day.Among them,the internal knuckle pattern based on hand images has the advantages of clear texture,obvious lines,less easy to wear and forge,and stable features compared to fingerprints and palm prints.However,many hand images are collected from cell phones in an open environment,and these images are affected by uneven lighting,scale changes,blurred or missing knuckle edges due to finger-joining,and complex backgrounds,resulting in low accuracy and poor stability of internal knuckle pattern recognition under complex links.To address the above problems,this project takes the internal knuckle patterns of hands captured by cell phones under indoor illumination and black background as the research object,and firstly constructs a library of hand images captured by cell phones under open environment,then extracts the internal knuckle pattern samples through a series of pre-processing processes,studies the internal knuckle pattern recognition method with fusion feature coding,and finally conducts a large number of experiments on the constructed internal knuckle pattern database for the relevant methods of the subject research.Finally,a large number of experimental validation analyses were conducted on the constructed internal fingerprint database for the relevant methods of the subject research.The main research work of this thesis is as follows:(1)Internal knuckle pattern image acquisition and pre-processing.The hand images are first acquired by the cell phone camera,and then preprocessed.The hand preprocessing steps include:bin arization,morphological processing,contour extraction,localization and separation of knuckles,ROI image extraction of internal knuckle patterns,image normalization,image enhancement,etc.Among them,the gradient algorithm is used in inner knuckle pattern ROI image extraction,which can better separate the inner knuckle pattern ROI image;adaptive histogram equalization is used in inner knuckle pattern ROI image enhancement,which makes the pattern information more prominent.Finally,an expanded dataset is used for the inner knuckle pattern ROI images to prevent overfitting problems in subsequent experiments.(2)The hand images captured by the cell phone camera will have fingers together,which will lead to blurred or missing edges of the separated knuckles.To address this problem,a convolution-based stencil edge detection algorithm is designed for finger localization and knuckle separation,in which a fully connected neural network is designed to perform error boundary fitting for knuckles with blurred or missing edges.The experimental results show that this method can repair the error boundary to obtain the complete finger boundary and effectively separate the inner knuckle pattern ROI images.(3)An improved feature coding-based internal knuckle pattern recognition method is designed.This topic is based on the CompNet feature coding network,which combines the traditional filter extraction and convolutional neural network extraction methods and is more resistant to light.In this thesis,we improve the convolutional neural network part of CompNetpost-processing unit(PPU),add MobileNetV2 inverse residual structure and coordinate attention module to its basic network structure,and take into account the position relationship based on the channel attention.The rich directional ranking information of inner knuckle patterns is utilized to extract higher-level features.Through experimental comparison,the correct recognition rate of internal knuckle patterns based on the improved CompNet network(CN1)can reach 98.6%,which is more than 3%higher than that of the original CompNet algorithm.(4)An internal knuckle pattern recognition method based on two-dimensional matching distance fusion was designed.Based on the improved CompNet network(CN1),the differences and complementarities between different feature coding algorithms such as CompCode,OrdinalCode,RLOC and BOCV are fused in a fractional layer,and the multidimensional feature vectors constructed from multiple matching distances are classified and recognized using a support vector machine(SVM).Through experimental comparison,the improved CompNet(CN1)and BOCV algorithms for two-dimensional matching distance fusion recognition are the best,i.e.,the accuracy of the method in this thesis reaches 99.2%for the recognition of internal knuckle patterns collected from cell phones in indoor scenes.Therefore the subject has good research application value.