Manuscript received June 20, 2025; revised July 24, 2025; accepted September 29, 2025; published February 27, 2026.

Abstract—Detecting fake and real smiles are challenging due to the subtle differences in facial muscle movements, particularly around the eyes and mouth. This paper presents a new hybrid approach that integrates a Convolutional Neural Network (CNN) with Dense Scale-Invariant Feature Transform (SIFT) descriptors to improve the accuracy and robustness of fake smile detection. Unlike conventional CNNonly methods, which rely solely on automatically learned features, our approach combines deep learned spatial features with handcrafted local descriptors to capture finegrained expression variations. Facial images were downscaled to 224×224 resolution, preserving essential features while reducing computational cost. To address class imbalance and data scarcity, the dataset was expanded through up-sampling and augmentation to 17,000 images per class. The CNN branch consisted of three convolutional layers with increasing filter sizes, followed by batch normalization and dropout for regularization. In parallel, the Dense SIFT branch extracted local invariant features, which were passed through fully connected layers. Both feature sets were concatenated and classified using a final SoftMax layer. This fusion approach achieved a test accuracy of 98.8% over 40 epochs, demonstrating improved performance compared to CNN-only baselines. The results highlight the potential of combining handcrafted descriptors with deep learning for real-time applications in human–computer interaction and emotion recognition systems.

Keywords—Convolutional Neural Network (CNN), smile detection, Scale-Invariant Feature Transform (SIFT), handcrafted features, fake smile, real smile, facial expression recognition, classification

Cite: Reem K. Sahi, Noha A. Hikal, and Heba Kandil, "Fake and Real Smile Detection Using a CNNBased Model with Handcrafted Features Using Scale-Invariant Feature Transform," Journal of Image and Graphics, Vol. 14, No. 1, pp. 149-162, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC-BY-4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.