End-to-End Canine Emotion Recognition from Images to Video: EfficientNetV2-S and Visual Interpretability Study

This study presents a system for the automatic classification ofdoge motions-happiness, s adness, a nger, and relaxation-using the EfficientNetV2-S a rchitecture, t rained on images and adapted for video inference. The model was trained on the public Dog Emotion dataset from Kaggle, employing transfer learning with ImageNet1K_V1 weights, combined with data augmentation techniques and stratified cross-validation. In addition, Grad-CAM was integrated as a visual explainability tool, enabling the identification of r elevant a natomical regions associated with each emotion. The system achieved an 85.43% accuracy on the test set, representing the best performance reported to date on this dataset, surpassing the results presented in previous state-of-theart studies. These results, combined with strong consistency in the visual activations, validate the potential of this approach as an efficient and explainable tool for canine emotional analysis in real-world contexts.