Enhancing human-dog interaction through deep learning and explainable AI

This article ventures into the novel field of recognizing dogs' emotions, employing the Deep Learning and Transfer Learning techniques. A unique aspect of our study was the creation of a robust dataset from scratch, based on the latest research in the field of dogs' emotion recognition. Using this dataset, we trained a model employing a transfer learning approach, fine-tuning established architectures to accurately identify dogs' emotions. Our results demonstrate substantial success rates, underscoring the efficacy of these methodologies in enhancing human-dog interaction and improving animal welfare. To ensure that our model's predictions were transparent and interpretable, we incorporated an eXplainable Artificial Intelligence (XAI) approach, utilizing Gradient-weighted Class Activation Mapping (Grad-CAM). This technique offers visual elucidations of the model's predictions, identifying the critical regions within the images for the emotional classification. Our findings underscore the substantial potential of deep learning and transfer learning approaches for understanding canine emotions. We evaluated ten image classification models, including convolutional neural network architectures, transformer-based models (ConvNeXt), and a classical machine learning approach based on Support Vector Machines (SVM) using DINO v2 features. The highest performance was achieved by the YOLO v11 (You Only Look Once) architecture, with both accuracy and F1-score reaching 0.84. Notably, the DINO v2-SVM model ranked second, attaining an accuracy of 0.75 and an F1-score of 0.76, while the ensemble model combining MobileNet, EfficientNet, and ResNet50 (Residual Network with 50 layers) achieved third place, with both metrics equal to 0.75.