<![CDATA[Dog breed classification remains a challenging task in computer vision due to high inter-class visual similarity, pose variations, changes in illumination, and complex background conditions. Conventional convolutional neural network (CNN) approaches often struggle to capture global contextual dependencies and subtle discriminative features. This study proposes a hybrid deep learning framework that integrates YOLOv8n for object detection with the Vision Transformer (ViT-B/16) for dog breed classification. The dataset comprises 14,181 dog images collected from the Tsinghua Dogs Dataset and supplementary real-world sources, spanning 10 dog breed categories. The proposed framework includes image preprocessing, data augmentation, transfer learning, and Bayesian hyperparameter optimization using Optuna to enhance model generalization. YOLOv8n is employed to localize dog regions, which are subsequently resized and passed to the Vision Transformer for global feature representation learning. The model is evaluated on 2,133 unseen test images. Experimental results demonstrate that the proposed framework achieves an accuracy of 97.98% with macro and weighted F1-score values of 98.76% and 97.98%, respectively. Comparative experiments against standalone ViT-B/16 and EfficientNetV2M architectures futher confirm the effectiveness of the proposed hybrid YOLOv8n–ViT-B/16 framework for dog breed classification.]]>
Copyrights © 2026
