Authors:
Mykhailo Paslavskyi, Ibrahim Ahmed Ibrahim Abd Elaal El Sayed, Lin Chih-Hao, Amarilys González García, Vinothini Kasinathan, Kohila Malar Kalesamy
Addresses:
Department of Computer Science, Ukrainian National Forestry University, Lviv, Lviv Oblast, Ukraine. Department of Biomedical Sciences, Zewail City of Science and Technology, October Gardens, Giza Governorate, Egypt. Department of Manufacturing, Taipei Tech Industries, Da'an District, Taipei, Taiwan. Department of Research and Development, Placental Histotherapy Center, Playa, Havana, Cuba. School of Computing, Asia Pacific University of Technology and Innovation, Kuala Lumpur, Wilayah Persekutuan Kuala Lumpur, Malaysia.
Abstract:
In computer vision, classifying dog breeds is a difficult problem because there are many breed classes, high similarity among related breeds, variations in pose and illumination conditions, and the presence of noisy, low-quality, or blurry images. In this paper, researchers propose a transfer learning approach using TensorFlow Hub to develop a large-scale multi-class image classification system for accurate dog breed identification. Researchers use pre-trained architectures like MobileNetV2 and EfficientNet as feature extractors since they can build rich, discriminative image representations at lower computational cost than training a deep neural network. The Kaggle Dog Breed Identification dataset has over 20,000 labelled photos of 120 dog breeds, which researchers used to train, validate, and assess the model. Image quality, data diversity, and overfitting are improved by resizing, pixel normalisation, and data augmentation through rotation, flipping, zooming, and shifting. The classification model is trained with categorical cross-entropy loss and an optimisation approach and assessed using Top-1 and Top-5 accuracy. Experimental studies reveal that transfer learning can classify visually related breeds well. The study also addresses class imbalance, intra-breed differences, background interference, and misclassification among breeds with similar physical traits. The proposed method shows that transfer learning can be used for large-scale dog breed recognition in pet identification systems, veterinary support, animal shelters, rescue services, mobile apps, edge devices, and real-time intelligent animal monitoring systems.
Keywords: Transfer Learning; Breed Recognition; Deep Learning (DL); Computer Vision; Data Augmentation; TensorFlow Hub; Class Imbalance; Background Interference; Veterinary Support.
Received on: 21/08/2025, Revised on: 10/10/2025, Accepted on: 25/10/2025, Published on: 07/06/2026
DOI: 10.64091/ATICR.2026.000341
AVE Trends in Intelligent Computing Research, 2026 Vol. 1 No. 2 , Pages: 91-101