Abstract:
The domain of facial feature identification (FER) has garnered significant attention in the fields of computer vision and pattern recognition, with deep learning approaches demonstrating their effectiveness ever since their inception. These solutions, however,
need a large number of parameters, making it impossible to use them for practical applications on embedded systems with limited resources. To address the limitations of deep learning-based FER models, we propose an efficient convolutional neural network (CNN) architecture capable of real-time facial expression recognition. Our model is specifically designed for deployment on
embedded platforms, aiming to mitigate the aforementioned drawbacks associated with existing deep learning-inspired FER structures. By keeping an appropriate balance between accuracy and computing performance, the CNN version that was built provides the most reliable overall performance. By utilizing the FER benchmark dataset (FER2013), we thoroughly assessed the performance of our CNN model, taking into account crucial evaluation parameters including recognition accuracy, precision, recall, and F1-score. Following that, we further enhanced the CNN model by leveraging the TensorRT SDK, resulting in an optimized
version that offers a portable solution characterized by efficient inference and high throughput. The FER system we developed demonstrated remarkable effectiveness by achieving competitive accuracy while significantly improving the execution speed by multiple folds was demonstrated by comparison analysis findings with other cutting-edge methods. Similar examination results with the other better strategies uncovered the viability of the planned FER framework, The FER system attained impressive accuracy while simultaneously achieving a substantial multiple-fold enhancement in execution speed.
