Efficient Facial Expression Recognition Through Lightweight CNN Technique on Public Datasets

The exploration of sentiments through facial expressions is a captivating domain with applications across security, healthcare, and human–computer interaction, where understanding sentiments is primarily about interpreting an individual's stance from a piece of text. However, in the context of non-verbal communication, it extends to the interpretation of emotions conveyed through facial expressions. This research endeavor aims to push the boundaries of machine-assisted sentiment predictions by refining models and methods for more accurate emotion recognition. A major obstacle in this pursuit is accurately identifying emotions from images not constrained by controlled conditions, often marred by poor visibility, shadows, or inconsistent lighting. Our study presents an innovative approach employing a lightweight convolutional neural network technique. This technique addresses various challenges to enhance the accuracy and reliability of sentiment analysis after incorporating pre-processing techniques such as sharpening to deal with image inconsistencies and histogram equalization to manage contrast variations, which improve image quality but include some artifacts in the image that are further resolved through the application of contrast limited adaptive histogram equalization. The technique’s effectiveness is underscored by its performance, achieving a 52% accuracy rate on raw, unconstrained images of FER-2013 and a 70% accuracy rate following the application of our pre-processing technique to the same dataset. The demonstrated proposed lightweight model not only performed exceptionally well over FER-2013 but also over CK+, RAF-DB, KDFE with an accuracy of 99.2%, 84.4%, and 94%. Furthermore, the proposed technique shows strong performance on real-time images captured via webcam.