HUMAN IDENTIFICATION BASED ON BIOMETRIC RADIOGRAPHS: A FORENSIC APPROACH

This invention presents holistic, multimodalrhetoricalhuman identification techniques victimization decision-level fusion for hand and dental radiographs. The projected human identification system demonstrates the novel application of assorted existing texture descriptors for the feature extraction of hand and dental radiographs. We've got used grey level Co-occurrence Matrix (GLCM), Kirsch's Filter, native Binary Pattern (LBP), changed native Binary Pattern (MLBP), Double native Binary Pattern (DLBP), Transition native Binary Pattern (t-LBP), native Ternary Pattern (LTP), twin Cross Pattern (DCP) and Gabor remodel (GT) for the feel feature extraction. Further, we've got planned a changed ALTP methodology that tackles the matter of corruption of central pixel within the native window because of noise by considering the native window's average for the LTP's computation. This methodology slightly outperforms the baseline LTP and former texture-based approaches. For classification, we've got used the K-Nearest Neighbour Classifier (KNN), Feed-forward Neural Network (FNN), Classification Tree (CT), and multiple Support Vectors Machine (m-SVM) at the side of every type of option extracted from radiographs. The proposed ALTP has given improved human identification accuracy (84.40%) compared to traditional texture-based techniques such as GLCM (59.20%), Kirsch's filter (63.20%), LBP (77.60%), MLBP (79.20%), DLBP (81.60%), t-LBP (82.40%) and LTP (83.20%) using m-SVM classifier. Further, we have investigated the Histogram of Oriented Gradients (HOG) for the hand and dental radiograph's shape description. It is observed that shape features outperformed the texture features for the collected dataset and resulted in the accuracy of 92.80% compared to promising texture description techniques such as ALTP (84.40%), DCP (87.20%), and GT (87.20%). Traditional texture and shape based techniques result in poor feature representation, lower-order correlation, and low discrimination power of raw features. We have developed three-layered deep convolutional neural network (DCNN) which m-SVM resulted in a 99.60% recognition rate for human identification using collaborative hand and dental radiographs. For the final DCNN implementation, we have selected a 3x3 kernel, six filter kernels, a 2x2 maximum pooling window and the 2-pixel stride. DESCRIPTIVE DRAWINGS: Hand/ nta\ RaiahcImage s ConvolutionalINeural NewrkCN I-ers ClsiiainTeIedfradM lil upr Neighbor~~~ ~ ~ ~ ~ ClsiirIN) Nua