posted on 2022-07-26, 12:58authored byAhmad Ghaffari
Application of iris recognition for human identification has significant potential for developing a robust identification system considering that the iris pattern of individuals is unique and is even differentiable from left to right eye and is almost stable over the time. This PhD research aims at developing novel iris segmentation and recognition techniques. Research show that the performance of iris recognition methods depend on the accuracy of their input iris images. However, the performance of existing traditional wavelet-based iris segmentation techniques is limited as the wavelet transform is shift variant. In this research the non-decimated wavelet transforms, which is a shift invariant transform, and Fuzzy reasoning are utilized to develop an iris segmentation algorithm called Non-Decimated Wavelet and Fuzzy based Iris Segmentation (NDWFIS). Experimental results on images of the benchmark image datasets show that the proposed method segments iris image more accurately, improving the recognition accuracy of any consequent iris recognition method. The proposed NDWFIS algorithm employs a histogram-based method to select a pixel within the pupil area of the input image and creates several radii vectors with respect to this pixel. It then performs a non-decimated wavelet transform on each resulting radii vector, extracting its wavelet sub bands. The transition from iris to limbus pixels generates a positive edge in wavelet high frequency sub bands. Hence, the proposed NDWFIS algorithm sets all negative wavelet coefficients to zeros. The resulting wavelet sub bands’ coefficients are normalised and fuzzified to reduce the effect of noise on detection of strongest edge in the radii vector. The information across fuzzified wavelet sub bands’ coefficients is combined using Fuzzy interstation. The coefficient with the highest Fuzzy membership value in the resulting combined vector is illustrate the location of the strongest edge. The extracted edges within the radii vectors are interpolated in a 2D matrix to create the outer boundary of the iris. A histogram algorithm is then applied on the pixels within the extracted region to determine the interior iris boundary. The performance of the proposed method on synthetic noisy step signal shows the robustness of the method in accurately locating the edge at highly contaminated step signal. Traditional wavelet-based iris recognition methods generate superior performance to their anchor image-based methods due to the time-frequency localisation property of wavelet transform. However, time-frequency localisation of the wavelet transforms is limited to horizontal and vertical directions. The aim of this research is improving the performance of the iris recognition systems using Curvelet transform. Since the Curvelet transform can extracts time-frequency information of the iris image from more than two angles that the wavelet transform does. Based on this assumption two iris recognition method named: I) an Intensity Separation Curvelet based PCA (ISC-PCA)
4
iris recognition method and II) An Intensity Separation Curvelet based Iris Recognition using Grey Level Run Length Algorithm (ISCIR-GLRL) are presented in this PhD theists. The proposed ISC-PCA method first extracts a rectangularize iris image from the input eye image using Daugman Rubber Sheet method. It then applies Curvelet transform on the resulting rectangularize iris image, divides it into its curvelet sub bands. The resulting complex coefficients within the sub bands of the same level are concatenated to form two frames. Each resulting frames’ coefficients are normalised separately and split into several bands based on their values. The coefficients in the resulting bands of each frame are then vectorized and concatenated, forming a 2D matrix. It then performs the Conventional PCA on the resulting 2D matrix computing its eigenvectors, which are used as feature for iris recognition. Experimental results on the images of CASIA-Iris-Interval benchmark eye image dataset shows the merit of the proposed method to its anchor PCA based methods. The proposed ISCIR-GLRL method takes a rectangularize iris image and applies a Curvelet transform on it, dividing the iris image into its curvelet sub bands. It then concatenates the coefficients in the resulting Curvelet sub bands, shaping a single matrix. It then uses the Grey level Run Length Matrix (GLRLM) algorithm to extract iris features. The resulting iris features are used to perform iris recognition. Experimental results on the image of CASIA-Iris-Interval benchmark eye image dataset, show that the proposed ISCIR-GLRL methods gives superior results to other curvelet based methods.