A study on lightweight facial expression recognition based on enhanced network models

This paper introduces a neural network designed to be lightweight (DL) for face expression recognition using dilation convolution without increasing parameters, enabling efficient global feature extraction. Incorporating Leaky Relu mitigates gradient vanishing and enhances training robustness. Experiments show a 71% and 52% reduction in parameters for DL_Vgg19 and DL_ResNet34, respectively, with minimal accuracy loss, validating the method's efficiency.