CNN-Driven Enhancement in Facial Emotion Recognition Systems

Abstract

Facial Emotion Recognition is crucial for developing systems used in human-computer interaction, mental health diagnostics, and security applications. Accurately detecting and  classifying human emotions from facial expressions is essential in these areas. This study aims to create a robust and high-accuracy model for recognizing and classifying human emotions from facial expressions using convolutional neural networks (CNNs). It emphasizes the use of transfer learning and data augmentation to improve the performance of the model on FER2013, which contains 35,887 black and white images representing seven emotions (surprise, happiness, sadness, disgust, anger, fear, and neutral). Each image is resized to 48x48 pixels to reduce computational complexity. The proposed model uses a pre-trained VGG16 CNN, fine-tuned with the FER2013 data. Preprocessing steps include resizing images to 224x224 pixels, normalization, and applying data augmentation techniques like rotation, scaling, and shifting to improve model robustness. The CNN analyzes features present in the input images and then runs them through fully connected layers to produce a distribution of probabilities across seven emotional categories using a softmax activation function. The algorithm is trained using the categorical cross-entropy loss function and the Adam optimizer, initializing with a learning rate of 0.001. This rate decreases tenfold every ten epochs. The batch size is 32, and training lasts for 50 epochs with early stopping. The use of transfer learning and data augmentation significantly improved the model's accuracy, achieving 83% on the test set. Performance measures like accuracy, completeness, and the F1-score were computed for each emotion category with the 'happy' class showing the highest precision (0.92) and recall (0.93). The confusion matrix revealed a high true positive rate and identified areas of misclassification, particularly between similar emotions like sadness and neutral. The training and validation loss curves indicated a good fit without overfitting, as shown by the close convergence of these values. Throughout the training period, the level of precision consistently rose, leveling off at approximately 83% for both the training and validation datasets, indicating a steady improvement in accuracy. The model shows potential for applications in mental health monitoring, customer feedback analysis, and personalized content delivery. Future work will focus on refining the model and applying it to real-world scenarios to enhance its robustness and accuracy in diverse environments.