Next-Gen dog breed identification using hybrid and custom CNN models

Dog breed identification using deep learning has become a key area of research in computer vision, with applications in automated pet recognition systems. This paper compares various deep learning models, including ResNet101, InceptionV3, Xception, and a hybrid model combining InceptionV3 and Xception, for dog breed classification. The Stanford Dog Breed dataset, consisting of images from 120 breeds, is used for training, focusing on five specific breeds to reduce computational time. Transfer learning is applied to fine-tune pre-trained models. The results show that the hybrid InceptionV3-Xception model achieves an accuracy of 94.27%, with a precision of 94.93%, recall of 94.07%, and an F-score of 94.42%. The custom CNN model, an extension to the proposed methods, outperforms the hybrid model, achieving a remarkable accuracy of 99.48%, precision of 99.49%, recall of 99.41%, and an F-score of 99.44%. These findings demonstrate that custom CNN architectures provide superior accuracy and efficiency compared to pre-trained models for dog breed identification.