InceptionNeXt Network with Relative Position Information for Microexpression Recognition

Abstract

Micro-expression recognition is one of the most challenging tasks in affective computing, aiming to identify subtle facial movements that are difficult for humans to perceive within a short period. In recent years, convolutional neural networks (CNNs) have been widely employed for micro-expression recognition. Despite achieving high recognition accuracy, these approaches still exhibit some limitations. Traditional CNNs learn solely from complete images, failing to accurately capture the fine details of facial micro-expressions. Traditional convolution uses small convolution kernels, which ignore global image features during feature extraction, affecting the accuracy of micro expression recognition. In response to the shortcomings of convolutional kernels, this paper proposes a large kernel convolutional neural network for micro expression recognition. Specifically: 1) Introducing a large-kernel convolutional neural network (LKCNN) that extends the model’s receptive field, enabling it to capture data relationships over a larger range. This enhances the model’s performance. To balance computational costs, the network utilizes Inception depthwise convolution, which decomposes expensive depthwise convolutions into three convolutional branches with small kernel sizes, along with an identity mapping branch. 2) The input is transformed from entire images to image blocks, followed by the incorporation of relative position attention (RPI). This not only magnifies detailed features but also filters important feature information, ensuring the full utilization of subtle information. Experiments were conducted on three publicly available databases: CASME II, SMIC, and SAMM. The results demonstrate superior performance of the proposed method in three-class and five-class recognition compared to state-of-the-art approaches.