Performance Evaluation of Complex Face Emotion Recognition Using Convolution Neural Networks Under Diverse Conditions

Facial emotion recognition methods are computer vision-based algorithms for identifying facial emotions from an input image or video, with several applications in surveillance systems, healthcare, education, and retail. A complex facial emotion can be visualized as a combination of two or more basic emotions. This study presents a model for recognizing complex facial emotions using a convolutional neural network, trained on the FER2013 dataset and tested on the CEED dataset. This study investigates the performance of the model in variations, such as illumination conditions, demographic diversity, and image characteristics. Systematic and controlled experiments were conducted to validate the performance of the model and calculate the Complex Emotion Scores (CES). The average CES remained at 80% for dark images with an accuracy of approximately 89%. Female images tend to maintain slightly higher CES scores in light conditions than male ones. However, CES remained consistent on image file types, and the CPU utilization was found to be high for certain image formats.