FER-MOTION: Facial Expression Recognizer in Multi-resolution Images Using a Lightweight Large Receptive Residual Network

Abstract

Real-world applications challenge facial expression recognition systems to adapt to various input image resolutions. Specifically, two-stage methods that rely on face patches from face detection tend to produce limited information for low-resolution cases and slow in the inference stage for high-resolution input. Besides, a high-performance facial emotion vision-based system requires an adaptive deep learning model with low parameter usage and computational cost. This work proposes a novel facial emotion recognizer in multi-resolution input (FER-MOTION) with high performance and cost-efficiency. The proposed network offers a lightweight CNN approach that is improved from MobileNetV2, offering a large kernel receptive module and a pyramid enhancement module, each designed to improve effectiveness and efficiency. This approach introduces a new extractor module capable of discriminating facial emotion features in a lightweight operation by capturing a larger spatial area at each network stage. A group-based attention module involving a pyramid spatial map is proposed to overcome the saturation performance of the extraction network. Comprehensive experimental results demonstrate that the proposed CNN architecture achieves high accuracy across varying image resolutions. The experiment is conducted on three benchmark facial expression datasets: KDEF, RAF-DB, and FERPlus. Analyses and comparisons of computational and parameter efficiency show that the proposed model is 3.8 times lighter in parameters and 1.8 times more efficient in floating-point operations than MobileNetV2.