Image processing-based identification and recognition in cattle using retinal biometrics

This thesis aims to develop a cattle retina-based identification and recognition system to address the harmful effects of conventional identification methods on cattle, such as infection, pain, and identity loss. A large dataset was collected and processed using both classical image processing and deep learning methods. Initially, 2,460 retina images were acquired from 300 cattle. Within the scope of classical methods, green channel selection, CLAHE histogram equalization, morphological operations, image subtraction, and noise filtering were applied. Fuzzy C-Means, K-Means, and Level-Set methods were employed for retinal vessel segmentation. The performance of BRISK, FAST, HARRIS, SIFT, and SURF techniques was evaluated for feature extraction. In the classical-method-based identification system, a combination of K-Means and SIFT achieved an accuracy of 95.6%, while in the recognition phase, the Level-Set + SIFT approach reached 87.9% accuracy. For deep learning-based methods, the U-Net, SA-UNet, and UNet++ models were trained using manually annotated images and data augmentation techniques. U-Net, in particular, stood out with a 97.4% validation accuracy in retinal vessel segmentation. Furthermore, the U-Net-based recognition system attained an 88.64% accuracy using the HARRIS feature extraction method. The results demonstrate that identification and recognition tasks can be successfully carried out using retinal images, and that deep learning-based approaches offer higher accuracy and reliability compared to classical methods. It is concluded that this approach can effectively prevent cattle theft and identity fraud, thereby reducing harm to the animals. Keywords: Retinal Identification, Animal Recognition, Deep Learning, Image Processing, SIFT, SURF, U-Net, SA-UNet, UNet++