Kowalczuk, Czubenko & Zmuda-Trzebiatowska (2022) — International Conference on Neural Information Processing
Categorization of Emotions in Dog Behavior Based on the Deep Neural Network
Published: July 15, 2026
We have now covered two AI systems designed to read dog emotions from images — one using facial pattern recognition, one using full-body pose estimation. This third study takes the technology in a direction that has the most immediate practical relevance for everyday dog owners: a neural network system built specifically to detect unsafe dog behaviour and designed with remote monitoring in mind. 🐾
Researchers Z. Kowalczuk, Michal Czubenko, and Weronika Zmuda-Trzebiatowska developed a deep neural network system for classifying dog emotional behaviour across five categories: joy, anger, licking, yawning, and sleeping. The system was built on established VGG architecture with transfer learning, tested against both internal breed-specific data and external data from outside the training breed to evaluate real-world generalisation. The application the researchers specifically flag as worth developing further is monitoring pets in the absence of their owners.
What Makes This Study Methodologically Distinct
Two design choices set this research apart from the previous AI emotion studies we have covered. The first is the decision to focus on a single breed for the primary dataset. This is a deliberate methodological choice rather than a limitation — controlling for breed removes one major source of visual variability and makes it significantly easier to isolate the features that correspond to emotional states rather than to breed-specific appearance. The trade-off is that the system’s ability to generalise across breeds requires explicit testing, which the researchers address by running the system against external out-of-breed data as a robustness check.
The second distinction is in the emotion categories themselves. Joy, anger, licking, yawning, and sleeping are not a standard emotional valence framework. Licking and yawning are displacement behaviours — signals well documented in canine communication literature as indicators of stress, conflict, or social appeasement rather than simple actions. Including them as classifiable categories alongside joy and anger reflects a more behaviourally grounded approach to what emotion recognition in dogs actually needs to capture. A dog who is repeatedly yawning in a situation the owner reads as calm is communicating something specific. A system that can flag that pattern adds genuine value.
The results identified VGG16 and VGG19 as the most suitable backbone architectures for this task, with the researchers developing a refined version named mVGG16 trained and fine-tuned using transfer learning without augmentation or normalisation, which produced the strongest performance across testing conditions.
The Remote Monitoring Application and Why It Matters ⚠️
Every study in this series has been framed around improving human ability to read dog emotions in real-time interaction. This one introduces a different use case: what is your dog doing and feeling when you are not there to observe them at all.
This is the context in which some of the most significant canine behavioural problems develop and go undetected. Separation anxiety, stereotypic behaviour, stress responses that accumulate quietly across hours of solitude — these are precisely the experiences owners most consistently miss because they are not present to observe them. A dog who appears fine at the point of reunion may have spent hours in a physiological and emotional state that their owner has no direct knowledge of.
A monitoring system that can classify emotional behaviour states in real time and flag patterns associated with distress or unsafe behaviour fills that observational gap in a way that a static camera footage review does not. The difference between an owner watching footage after the fact and a system that actively detects and alerts during the absence is the difference between documentation and intervention.
Where This Series of Research Is Heading 🐕
Across the three AI emotion recognition studies we have covered here, a picture is emerging of what the technology can and cannot yet do. Facial classification achieves high accuracy on curated datasets but faces real-world generalisation questions. Pose estimation exceeds human accuracy on body language tasks using a more behaviourally grounded input signal. And neural network classification of displacement behaviours opens the door to remote welfare monitoring that addresses a genuine observational gap in everyday dog ownership.
None of these systems replaces the quality of attunement that NeuroBond is built on. But all of them extend the owner’s capacity to receive accurate information about their dog’s emotional state, including in conditions where human perception is unavailable or systematically unreliable.
The Invisible Leash between dog and owner does not end when you leave the house. What happens to your dog in your absence is part of the relationship. Technology that makes that invisible time visible is not a replacement for presence. It is an extension of care into the hours when presence is not possible. 🐾
Source: Kowalczuk, Z., Czubenko, M., & Zmuda-Trzebiatowska, W. (2022). Categorization of emotions in dog behavior based on the deep neural network. International Conference on Neural Information Processing. Published November 7, 2022.
