An Enhanced Lightweight Framework with Selective Kernel and Attention for FER using Modified Bottleneck

Facial Expression Recognition (FER) plays an important role in enabling machines to interpret human emotions. While recent advances have led to high-performing models and many of these remain too computationally demanding for use on resource-limited devices. In this work, we introduce a compact yet effective FER model based on modified inverted residual blocks, augmented with two well known key modules: the Coordinate Attention (CA) for enhance spatial attention by encoding directional relationships across facial regions and a depthwise convolution-based Selective Kernel (SK) module for dynamically adjusts receptive fields by selecting from multiple kernel sizes. We have evaluated on the RAF-DB and FER2013 benchmarks, it achieves accuracies of 86.83 % and 70.25 %, respectively, with a model size of only 1.17 million parameters compared to the lightweight state-of-the-art-models. These results highlight the model's ability to making it well-suited for deployment on edge devices.