<![CDATA[Facial Expression Recognition (FER) remains a challenging problem in computer vision, particularly under subject-independent conditions in which models must generalize to individuals not seen during training. This study reports a controlled comparative evaluation of three Convolutional Neural Network (CNN) architectures — MobileNetV3-Large, EfficientNet-B3, and ResNet50 — using the Extended Cohn-Kanade (CK+) dataset (981 apex-frame images, 118 subjects, seven emotion classes). All models were trained and tested under identical experimental conditions with a subject-disjoint partition (72/23/23 subjects for training, validation, and testing), so that observed performance differences may be attributed primarily to architectural design. The results indicate that MobileNetV3-Large attains the highest test accuracy of 95.16%, exceeding EfficientNet-B3 (93.01%) and ResNet50 (91.94%), while requiring the fewest parameters (~5.4M) and the shortest inference latency (~8.2 ms per image). A multi-dimensional evaluation covering per-class metrics and computational cost is also reported. These observations provide preliminary architectural guidance for FER deployment in resource-constrained environments; however, because they are derived from a single dataset and a single subject split, broader claims should be confirmed on more diverse benchmarks.]]>
Copyrights © 2026
