A Generalizable Facial Emotion Recognition Tool: Cross-Dataset Validation of a Haar Cascade - CNN Pipeline

Most Facial Emotion Recognition systems are built and evaluated on a single dataset, so their reported accuracy says little about how they would perform on genuinely new data. Our contribution is not a new architecture but a new way of validating one: rather than reporting accuracy from a single train/test split, we train a Haar Cascade--CNN pipeline exactly once and then test it, without any retraining or modification, on a second, independently collected dataset -- in both transfer directions -- alongside a mixed-source configuration. This directly demonstrates, rather than assumes, that the tool generalizes to unseen data. To enable this, we built a custom, web-crawled dataset (Google Images and Freepik, seven emotion classes) that is intentionally more realistic and diverse than curated benchmarks, and paired it with a larger open-source dataset. Faces are localized via Haar Cascade, then classified using a compact CNN trained end-to-end with 5-fold cross-validation. The model reaches 86\% accuracy on the custom dataset and 41\% on the open-source dataset alone (79\% mixed). The key result is cross-dataset transfer: 65\% (custom$\to$open-source) and 40\% (open-source$\to$custom) accuracy -- both well above a comparable CNN's reported accuracy on a standard benchmark -- showing the pipeline retains real classification power on data it never saw during training.