Yazar "Saleh, Abbadullah .H Saleh" seçeneğine göre listele

Listeleniyor 1 - 1 / 1
  • Küçük Resim
    Öğe
    IRIS SEGMENTATION AND RECOGNITION BASED ON DEEP LEARNING IN THE PRESENCE OF DISEASES
    (2022-07) Saleh, Abbadullah .H Saleh
    Two main steps are involved in any iris recognition system: iris segmentation and iris recognition. A lot of iris segmentation and recognition systems have been introduced in recent decades. Too little research has focused on eye pathology cases and their effects on iris segmentation and recognition systems. In the current study, a new deep learning-based iris recognition system is introduced in the case of eye disease. A novel dynamic circular Hough transform algorithm is designed and implemented in the iris segmentation step. The transfer learning approach is used to apply three different deep learning models (GoogleNet, ResNet50, and ResNet101) through the recognition step. Three separate datasets are used. The first one is the Warsaw Bio-Base V1 collection which contains 684 iris images of people with various eye disorders. The second dataset is the Warsaw Bio-Base V2 dataset, which has 1793 iris scans with more complex eye cases and a larger number of photos. The third dataset is the CASIA V3 Interval Iris dataset, which has 2639 healthy iris photos. Experiments are conducted under different training and evaluation scenarios. During those scenarios, many training considerations are taken into account: three deep learning models, different splitting criteria; colored and grayscale iris images; segmented and original iris images, and transfer learning as two layers of training. MATLAB 2020a is used to build all the needed software. Besides that, like deep learning and image processing etc. some toolboxes are used. Many ways are used to check the accuracy of the result models, such as training accuracy, validation accuracy, test accuracy, confusion matrix, TPR, FNR, PPR, FDR, and training duration. The ground truth of iris segmentation is built, and the results indicate a low FPR of 0.79% and an FNR of 5.49% for the Warsaw Bio-Base V1 dataset. Results indicate that GoogleNet has low computational time in all cases, but lower performance compared to ResNet models. However, the best recognition accuracy in scenario No. (6), where iris recognition accuracy is 98.5% and 99% by ResNet50 and ResNet101 respectively for only Warsaw Bio-Base Version one and exact results when only CASIA V3 Interval Iris is utilized. Additionally, the ResNet50 achieved the greatest accuracy for the Warsaw V2 at 97.26 %. In contrast, using mentioned datasets (Warsaw Plus CASIA), the ResNet50 achieved a 98% of iris test accuracy. The impact of eye diseases on iris segmentation and recognition is being investigated and evaluated. The findings revealed that eye diseases, in some cases, have a considerable impact on iris segmentation, particularly in the case of mixed diseases, pupil abnormalities, eye trauma, blindness, some retinal detachments, and bloody eye concerns. The results also show that several eye problems, such as cataracts, glaucoma, blurry conditions, some lens abnormalities, and some corneal problems, have no effect on iris segmentation when they exist separately. When it comes to iris recognition, eye illness has a smaller impact when it comes to iris segmentation. Some cases of blindness are easily recognized. The ocular scenario in which the iris is covered, or its structure is modified partially or wholly is the most impactful challenge to iris recognition. The results show that some unique circumstances can be easily incorporated into iris recognition systems. According to the results, some eye problems can make it hard for iris recognition to work. This should be checked out and fixed before using biometric systems.