Analysis and Discussion

The evaluation of the image classification model trained on a custom dataset using Google Teachable Machine reveals key insights into its performance and limitations. The model demonstrated strong accuracy and precision in classifying dog images, achieving a perfect precision score (100%) and high recall (94.12%), indicating both confident and consistent identification of this class. This outcome reflects the effectiveness of convolutional neural networks (CNNs) in capturing salient visual features, even within simplified architectures (LeCun, Bengio, & Hinton, 2015; He et al., 2016).

In contrast, the model showed lower performance for cat and bird images, with both classes yielding accuracy and recall around 70%, and bird precision at 78.57%. This performance gap may be attributed to intra-class variability, overlapping features, or less diverse representations in the training set—factors that are known to affect generalization in image recognition tasks (Mitchell, 2019). Additionally, the use of different image sources for training (Unsplash) and testing (Pexels) could have introduced a domain shift, impacting model accuracy due to inconsistencies in lighting, background, or composition (Google Teachable Machine, 2023; Pexels, n.d.; Unsplash, n.d.).

Despite these challenges, the model showed adequate generalization ability for a beginner-friendly framework. The architecture provided by Google Teachable Machine enables rapid deployment and experimentation, making it ideal for educational use and prototype development (Google Teachable Machine, 2023). However, performance on more complex categories like cats and birds could be improved through data augmentation, use of larger or more balanced datasets, and exploration of deeper or residual network architectures (He et al., 2016; Russell & Norvig, 2020).

Conclusion

This project successfully implemented a supervised image classification task using Google Teachable Machine to categorize animals into three classes: Cat, Dog, and Bird. The model demonstrated high accuracy and precision in classifying dogs, while its performance was more moderate for cats and birds—indicating areas for further improvement. These results highlight the importance of dataset diversity, consistent image sourcing, and the potential need for more advanced model customization (LeCun, Bengio, & Hinton, 2015; He et al., 2016).

To build on these results, future work will focus on expanding the model’s capabilities and practical relevance. This includes testing the model with real-time camera input to evaluate its performance in dynamic, real-world environments and experimenting with audio-based animal recognition, such as distinguishing between barks and meows. These enhancements aim to create a more robust and versatile recognition system, paving the way for broader applications in fields like pet monitoring, wildlife observation, and intelligent home systems (Mitchell, 2019; Mayer-Schönberger & Cukier, 2013).

The use of Google Teachable Machine proved effective for rapid prototyping and ease of implementation, especially for educational and lightweight applications (Google Teachable Machine, 2023). However, the relatively simple convolutional neural network (CNN) model used may benefit from further architectural enhancements and training with a more balanced and diverse dataset to boost generalization performance (Russell & Norvig, 2020).