This study evaluates an image classification model developed with data augmentation to identify mixed-breed dogs. While their popularity is rising, identifying mixed-breeds remains challenging due to overlapping parental traits and the purebred focus of existing systems. We trained a model on three mixed breeds—Maltese and Toy Poodle, Pomeranian and Toy Poodle, and Chihuahua and Dachshund—using 250 augmented images per breed, and visualized key focus areas using Grad-CAM.
The classification achieved high performance for Maltese and Toy Poodle and Chihuahua and Dachshund (recall: 1.00, precision: 0.89), but lower precision for Pomeranian and Toy Poodle (0.53), indicating a tendency for misclassification. Grad-CAM visualization confirmed that the model focused on distinct features depending on the breed, such as the face, ears, or facial markings (Figure 1).
Future work includes expanding training data and refining the architecture to realize a practical web-based system.
