In this paper, we propose a novel CNN-based model named as Deep Cross Feature Adaptive Network (DCFA-CNN) for facial expression recognition. The proposed DCFA-CNN model holds two major components: shape feature (ShFeat) block and texture feature (TexFeat) block, respectively. The ShFeat block is responsible to extract high-level responses, which leads to discriminate features from different expressive regions, while TexFeat block leads to hold micro/minute variations which defines structural differences in the expressive regions. Moreover, DCFA-CNN embedded a two-branch cross-relationship to collect information of ShFeat and TexFeat block. These different responses boost discriminability of the network by incorporating complementary features. The effectiveness of the proposed DCFA-CNN is evaluated extensively with four datasets: CK+, MUG, ISED and OULU-CASIA, over single-domain subject independent and cross-domain ethnicity independent experimental setups. The experimental results show a significant improvement of 21.8%, 21.55% and 6.43%, 17.9% as compared with MobileNet for 6- and 7-classes over ISED and OULU-CASIA. The extensive ablation experiments have done to validate the role of each module in DCFA-CNN framework.

在本文中，我们提出了一种新的基于 CNN 的模型，称为深度交叉特征自适应网络 (DCFA-CNN)，用于面部表情识别。提出的 DCFA-CNN 模型包含两个主要组件：形状特征（ShFeat）块和纹理特征（TexFeat）块，分别。ShFeat 块负责提取高级响应，这导致区分来自不同表达区域的特征，而 TexFeat 块导致保持微/分钟变化，这定义了表达区域中的结构差异。此外，DCFA-CNN 嵌入了一个两分支交叉关系来收集 ShFeat 和 TexFeat 块的信息。这些不同的响应通过结合互补特征提高了网络的可辨别性。所提出的 DCFA-CNN 的有效性通过四个数据集进行了广泛评估：CK+、MUG、ISED 和 OULU-CASIA，在单领域主题独立和跨领域种族独立实验设置上。实验结果表明，在ISED和OULU-CASIA的6-和7-classes上，与MobileNet相比分别提高了21.8%、21.55%和6.43%、17.9%。已经进行了广泛的消融实验以验证每个模块在 DCFA-CNN 框架中的作用。