Talk · International Conference on Information and Education Technology · April 17, 2026

Quantifying Model Uncertainty in Fine-Grained Image Classification for Veterinary Training Applications

Dina Koishiyeva, Anel Aidos, Dinargul Mukhammejanova, Assel Mukasheva, Jeong Won Kang

Abstract

Horse breed recognition was studied as a finegrained image classification task for veterinary education using the Horse Breeds dataset with seven classes. Six convolutional backbones were compared in a unified transfer learning setting with fixed feature extractors and the same classification head: MobileNetV2, EfficientNetB0, ResNet50, VGG16, DenseNet121, and Xception. Prediction reliability was assessed from softmax probabilities using entropy, less confidence, and probability margin, and the models were tested under Gaussian noise with standard deviation 25 and Gaussian blur with a 15 by 15 kernel. ResNet50 had low uncertainty values among the better performing models, with entropy 0.5634, least confidence 0.1936, and marginbased uncertainty 0.3038, and kept noise accuracy at 0.8657. EfficientNetB0 reached the highest blur accuracy at 0.7388 with intermediate uncertainty values, entropy 0.7536 and least confidence 0.2491. VGG16 showed very low uncertainty scores, entropy 0.2633 and least confidence 0.0971, despite lower macro F1. The combined analysis of accuracy, uncertainty, and robustness supports model selection for a prototype educational tool for horse breed recognition.

Advanced Neural Network ApplicationsDomain Adaptation and Few-Shot LearningSmart Agriculture and AImodel uncertainty quantificationfine-grained image classificationveterinary education

Quantifying Model Uncertainty in Fine-Grained Image Classification for Veterinary Training Applications

Abstract

More by Assel Mukasheva