Related Experiment Video
Updated: Jul 23, 2026
08:00
Decoding Natural Behavior from Neuroethological Embedding
Published on: October 3, 2025
Enhancing human-dog interaction through deep learning and explainable AI
Michał Kopczyński1, Michał Czubenko2
1Department of Decision Systems and Robotics, Faculty of Electronics Telecommunications and Informatics, Gdańsk University of Technology, Narutowicza 11/12, 80-233, Gdańsk, Pomeranian, Poland.
Scientific Reports
|July 1, 2026
Summary
This study introduces a new method for recognizing dog emotions using deep learning and transfer learning. The developed model significantly enhances understanding of canine emotions, improving human-dog interactions and animal welfare.
Area of Science:
- Artificial Intelligence
- Animal Behavior
- Computer Vision
Background:
- Understanding canine emotions is crucial for animal welfare and human-dog interaction.
- Existing methods for emotion recognition in dogs are limited.
- Novel approaches are needed to accurately interpret canine emotional states.
Purpose of the Study:
- To develop and evaluate deep learning and transfer learning models for accurate dog emotion recognition.
- To create a novel, robust dataset for training canine emotion recognition models.
- To enhance the interpretability of emotion recognition models using eXplainable Artificial Intelligence (XAI).
Main Methods:
- A custom dataset of dog emotions was created based on current research.
- Transfer learning was applied, fine-tuning established deep learning architectures.
- Ten image classification models were evaluated, including CNNs, transformer-based models (ConvNeXt), and Support Vector Machines (SVM) with DINO v2 features.
- eXplainable Artificial Intelligence (XAI) techniques, specifically Gradient-weighted Class Activation Mapping (Grad-CAM), were used for model interpretability.
Main Results:
- The YOLO v11 architecture achieved the highest performance, with accuracy and F1-score of 0.84.
- The DINO v2-SVM model demonstrated strong performance, ranking second with an accuracy of 0.75 and F1-score of 0.76.
- An ensemble model combining MobileNet, EfficientNet, and ResNet50 achieved third place with accuracy and F1-score of 0.75.
Conclusions:
- Deep learning and transfer learning approaches show significant potential for understanding canine emotions.
- The developed models can enhance human-dog interaction and contribute to improved animal welfare.
- XAI methods like Grad-CAM provide valuable visual explanations for model predictions, increasing trust and transparency.