Children's Expression Recognition Based on Multi-Scale Asymmetric Convolutional Neural Network

This paper proposes a multi-scale asymmetric convolutional neural network (MACNN), specifically designed to tackle the challenges encountered by traditional convolutional neural networks in the realm of children's facial expression recognition. MACNN addresses problems like low accuracy from facial expression changes, poor generalization across datasets, and inefficiency in traditional convolution operations. The model introduces a multi-scale convolution layer for capturing diverse features, enhancing feature extraction and recognition accuracy. Additionally, an asymmetric convolutional layer is integrated to learn directional features, improving robustness and generalization in facial expression analysis. Post-training, this layer can revert to a standard square convolutional layer, optimizing efficiency for child expression recognition. Experimental results indicate that the proposed algorithm achieves a recognition accuracy of 63.35% on a self-constructed children's expression dataset, under the configuration of a GPU Tesla P100 with 16GB video memory. This performance exceeds all comparative algorithms and maintains efficient recognition. Furthermore, the algorithm attains a recognition accuracy of 78.26% on the extensive natural environment expression dataset RAF-DB, highlighting its robustness, generalization capability, and potential for practical application.