针对传统卷积网络在处理部分相关性和特征冗余方面的不足导致表情识别精度不高的问题，提出了一种轻量化微表情识别方法。通过整合稀疏逆协方差估计（SICE）和空间通道重建卷积（SCConv）来提升图像分类性能。首先，通过将SICE作为新的网络层加入到ResNet50中，模型能够更有效地识别并利用特征间的部分相关性，这有助于降低由混杂变量引起的偏差。其次，SCConv的引入替换了标准卷积层，通过空间重建单元（SRU）和通道重建单元（CRU）优化模型结构。最后，实验结果表明，相比其它网络模型，基于稀疏逆协方差估计的表情识别网络在表情识别数据的分类精度达到90.75%，显著高于其它网络模型。研究的模型参数量与计算量分别为20.13M与3.61GFlops，相比其它模型，减少了模型的参数量与计算资源消耗。这表明稀疏逆协方差估计与空间通道卷积能够在表情图像分类任务中进一步提升分类性能。

To address the limitations of traditional convolutional networks in handling partial correlations and feature redundancy, leading to suboptimal facial expression recognition accuracy, this paper proposes a lightweight micro-expression recognition method. This method enhances image classification performance by integrating Sparse Inverse Covariance Estimation (SICE) and Spatial Channel Reconstruction Convolution (SCConv). Initially, incorporating SICE as a new network layer into ResNet50 enables the model to more effectively identify and utilize partial correlations among features, helping to reduce bias caused by confounding variables. Subsequently, the introduction of SCConv, replacing standard convolutional layers, optimizes the model structure through Spatial Reconstruction Units (SRU) and Channel Reconstruction Units (CRU). Finally, experimental results indicate that the facial expression recognition network based on sparse inverse covariance estimation achieved a classification accuracy of 90.75% on facial expression data, significantly outperforming other network models. The studied model demonstrated a reduction in both parameter count and computational cost, with values of 20.13M and 3.61GFlops, respectively, when compared to other models. This suggests that sparse inverse covariance estimation combined with spatial channel convolution can further enhance classification performance in facial expression image.

TP391???????

2021年重庆市属本科高校与中科院所属研究所合作项目：工业互联网内生安全关键技术研究与协同创新项目（HZ2021015）；重庆市教委科学技术研究计划重点项目（KJZD-K202305201）；重庆科技大学硕士研究生创新计划项目（ZNYKC2327）