ARCHIVES

Abstract: A lightweight CNN is designed for real-time detection of facial emotions, age, and gender. It integrates MTCNN for efficient face detection, passing coordinates to a custom emotion classifier and pre-trained Caffe models for age and gender prediction. MTCNN’s cascade detection optimizes memory and processing efficiency. The emotion model uses Global Average Pooling and depth-wise separable convolutions to improve interpretability and portability. Pre-trained Caffe models handle age and gender prediction with preprocessing like mean subtraction and blob formation. Real-time detection is achieved via OpenCV and Dlib, with results displayed when confidence exceeds 50%. Tested on the FER-2013 dataset, the emotion model achieves 67% accuracy using 0.496GB of memory, with a compact size of 872.9 kilobytes for deployment across both static and dynamic inputs.