The task of facial emotion recognition in resource-constrained scenarios requires a trade-off between recognition accuracy and model size. This paper

based on the verified real experimental results on the FER2013 dataset

organizes and analyzes the performance of lightweight facial expression recognition models. The content discussed in this paper includes: the preprocessed results of the FER2013 dataset

the one-time backbone comparison between MobileNetV3Small-FP32 and EfficientNetV2B0-FP32

and the FP32 and pruning experiments centered around MobileNetV3Small. The preprocessed FER2013 dataset contains a total of 35

887 images

including 28

709 training images and 7

178 test images

with 7 categories. In the backbone comparison

the Keras test accuracy of EfficientNetV2B0-FP32 is 0.6786

which is higher than that of MobileNetV3Small-FP32 at 0.6354; however

the latter has only 943

159 total parameters

corresponding to a TFLite FP32 file size of 3.5947MB

significantly smaller than the former's 5

928

279 parameters and 22.3811MB file size. In the mainline reproduction experiment

the Keras and TFLite FP32 test accuracies of MobileNetV3Small-FP32 are both 0.6290; under the current pruning configuration

the Keras and TFLite FP32 test accuracies of the stripped model are both approximately 0.56018

and the TFLite file size remains at 3.573MB

without demonstrating the benefit of reduced file size. The experimental results indicate that

under the current experimental settings

MobileNetV3Small can serve as a reasonable baseline backbone for lightweight research; however

the existing pruning paths have not simultaneously achieved accuracy maintenance and model size reduction. The formal analysis scope of this paper is limited to the experimental results that have formed a complete chain of evidence.