I trained 3 Convolutional Neural Networks (CNNs), including a basic CNN designed from scratch, ResNet-18 trained from scratch, and a version ResNet-18 pre-trained on ImageNet, to determine which CNN yields the most accurate and precise animal identification, on the ANIMAL-10N dataset consisting of 5 confusing pairs of animal classes.
After training, various experiments were performed on the models to learn about their abilities and the impacts of transfer learning on such tasks, using each neural network's top-1 classification accuracy, precision and recall. Its classification patterns were quantified using Confusion Matrices and its features were visualized and evaluated qualitatevely using Saliency Maps.