DEEPLNET: A LIGHTWEIGHT CONVOLUTIONAL NEURAL NETWORK FOR REAL-TIME FACIAL EMOTION RECOGNITION

Abstract

This paper introduces DeepLNet which is a small convolutional neural network that is capable of recognizing facial emotion in real-time with high accuracy and low computational cost. The model was trained and evaluated using a facial emotion dataset separated into 70 percent training, 15 percent validation and 15 percent testing data. DeepLNet recorded an overall accuracy of 92, exceeding those of traditional CNN (85) and models based on MobileNet (89). The evaluation measurements were good with precision of 91% and a recall of 90% and F1-score of 90.5, which indicated balanced and reliable classification. Analysis of the confusion matrix showed the highest accuracy in the recognition of happiness (94%), and surprise (93%), with slightly lower accuracy in fear (89%) and anger (88%). When it comes to computational efficiency, the proposed model used 2.8 million parameters versus 8.5 million in standard CNNs, and at the same time, it was faster than the traditional CNNs, with a higher processing speed of 35 frames per second (FPS), which is suitable in real-time applications. The performance under varying conditions revealed an accuracy of 93 percent in controlled setting, 88 percent under low-light conditions, and 86 percent in multi-face conditions. These findings indicate that DeepLNet is a good compromise between accuracy and efficiency and is therefore suitable in implementation on the resource-limited devices like mobile phones and embedded systems. The paper identifies the promise of lightweight deep learning models in the development of real-time emotion recognition applications.