Optimized Image Preprocessing for Enhanced Facial Emotion Recognition Using Deep Learning Architectures

Image recognition research of human emotions remains a vital complex task within social communication fields. Much better results are achieved in emotion recognition through deep learning (DL) when compared to conventional image processing approaches. “Facial emotion recognition” (FER) facilitates affective computing and efficient human-computer interaction. This work implements deep learning-based FER through the combination of FER-2013 dataset and DenseNet169 architecture to perform facial emotion recognition. Thirty-five thousand eight hundred eighty-seven grayscale images with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$48 \times 48$</tex> pixels comprise the dataset which implements CLAHE for enhancement and applies SMOTE to achieve balance. DenseNet169 functions as a feature extractor in the proposed deep learning model that conducts multi-class classification while employing ‘Global Average Pooling’ along with ‘fully connected layers’ followed by a softmax output. Given a 5fold cross-validation training approach the proposed model managed to achieve 97.71 % training precision and 91.96 % validation precision which exceeded performance levels of ResNet-50 (73.40%) and MobileNetV2 (58.00%). The model demonstrates effectiveness in performance analysis through accuracy results and loss trend analysis and confusion matrix evaluation and ROC curve interpretation yet shows misclassification in similar emotional categories. The DenseNet169 model shows robustness together with strong generalization abilities, although it exhibits small degrees of overfitting behavior, which makes it appropriate for practical FER deployment.