<p>Annotation of an image with numerous applications may be considered as one of the key challenges in computer vision. Deep neural networks like CNNs (Convolutional Neural Networks) model have proven highly effective in automatically discovering distinguishing features from raw pixel data. Conversely, algorithms based on instance-based learning, such as k-Nearest Neighbors (k-NNs), are recognized for their simplicity. The goal is to understand their strengths and weaknesses and propose a hybrid model that balances interpretability and accuracy. This study conducts an empirical analysis of several algorithms employed in supervised learning, including Support-Vector-Machine (SVM), and on CNN, specifically focusing on image annotation within a single-label multiclass annotation task. Based on these findings, a hybrid CNN-k-NN model is proposed, leveraging the best-performing CNN for feature extraction and k-NN for single-label annotation. On accuracy, CNN outperformed both SVM and k-NN, achieving over 90% on the dataset CIFAR-10. k-NN achieved lower accuracy (34–38%) due to its sensitivity to noisy data and high-dimensional feature spaces. The hybrid CNN-k-NN model achieved a balanced trade-off between accuracy and interpretability, reaching 78% overall accuracy with optimal k-values. The study demonstrates that while CNNs offer high accuracy, they lack interpretability and are computationally intensive, particularly with large datasets, which can lead to overfitting. In contrast, k-NN is simpler and more interpretable but faces accuracy challenges on complex datasets. By integrating the strengths of both methods, the hybrid CNN-k-NN model achieves a more favorable balance between interpretability and performance, thus presenting itself as a promising solution for image annotation tasks.</p>