REAL-TIME ASSESSMENT OF HUMAN EMOTIONAL STATE FROM FACIAL IMAGES USING THE EFFICIENTNET MODEL AND THE FACEEMOCDS DATASET

ТАТУ хабарлари · 2025-yil

Annotatsiya

The detection of emotional states from human facial expressions is a crucial research area in artificial intelligence and computer vision.This paper explores real-time emotion recognition using the EfficientNet-B0 model on the FaceEmocDS dataset.The FaceEmocDS dataset comprises 72,412 high-quality images, encompassing eight emotion classes: anger, contempt, disgust, fear, happiness, neutral, sadness, and surprise.Derived from multiple open-source datasets, it ensures diversity and balance after rigorous cleaning to remove lowquality or redundant samples.The model was trained via transfer learning, leveraging pre-trained weights from ImageNet, and fine-tuned over 30 epochs on the PyTorch framework.It achieved 74.32% accuracy on the validation set.To enhance generalization, data augmentation techniques were applied, including random rotations (up to 10 degrees), color jittering (adjusting brightness, contrast, and saturation), random resized cropping, horizontal flipping, and random erasing (with probability 0.7).Class weights were incorporated into the CrossEntropyLoss function to address imbalances, particularly in underrepresented classes like contempt and disgust.For real-time implementation, MediaPipe Face Detection was integrated for efficient face localization, followed by emotion classification with the EfficientNet-B0 model.The system operates at 20-30 frames per second (FPS) on standard hardware, making it suitable for dynamic environments.Test results yielded an accuracy of 73.18% and a weighted F1-score of 0.7327, with highest performance on happiness (93.08%) and lowest on contempt (63.05%).Confusion matrix analysis revealed common misclassifications, such as contempt with anger or disgust, highlighting areas for improvement.Compared to baseline models such as VGG and ResNet, this approach provides superior efficiency in terms of parameter count and inference speed while maintaining competitive accuracy.

Maqola ma’lumotlari
JurnalТАТУ хабарлари
Nashr sanasi2025-10-22
DOI10.61663/253tuitmct3

ТАТУ хабарлари jurnalidan boshqa maqolalar

ТАТУ хабарлари — barcha maqolalar