改进人脸特征矫正网络的遮挡人脸识别方法

现有人脸识别模型受口罩等遮挡因素影响导致准确率无法提升。当前主流研究方法将有无遮挡场景分开训练后，整合应用于多场景。针对遮挡人脸识别模型的局限性，提出一种改进人脸特征矫正网络（FFR-Net）模型。该模型可同时用于有无遮挡人脸识别并应用于口罩与眼镜遮挡两种识别场景中。人脸特征矫正网络模型提出了一种人脸特征矫正模块，为保证充分利用无遮挡区域特征信息，在该模块中的空间分支引入involution算子扩大图像信息交互区域，增强在空间范围内面部特征信息；在通道分支引入坐标注意力机制，捕获跨通道信息以增强特征表示，利于模型准确地定位识别目标区域；将Meta-ACON作为该模块新的动态激活函数，通过动态调整线性或非线性程度以提高模型泛化能力和计算准确度。最后，利用改进的人脸特征矫正网络模型在CASIA-Webface经处理的有无口罩遮挡人脸数据集上进行训练，其在LFW经处理的有无口罩遮挡数据集、Meglass数据集上的测试结果准确率分别达到了82.50%和89.75%，优于现有算法，验证了所提方法的有效性。

Improved face feature rectification network for occluded face recognition

The accuracy of existing face recognition models cannot improve due to the influence of masks and other occlusion factors. The current mainstream research methods integrate and apply the occluded and unoccluded scenes to multiple scenes after separate training. Aiming at the limitation of occluded face recognition model, this paper proposed an improved face feature rectification network(FFR-Net) model. This model could be used for face recognition with or without occlusion, and be applied to mask and glasses occlusion recognition scenes. FFR-Net proposed a face feature rectification module. In order to make full use of the feature information of the unocclusion area, the spatial branch of the module introduced involution operator to expand the image information interaction area and enhance the face feature information in the spatial range. The channel branch introduced coordinate attention to capture cross channel information to enhance the feature representation, which was conducive for the model to locate and identify the target area more accurately. Using Meta-ACON as a new dynamic activation function, it improved model generalization and calculation accuracy by dynamically adjusting the degree of linearity or nonlinearity. Finally, this paper trained the improved FFR-Net on the CASIA-Webface processed face dataset with or without mask occlusion. The accuracy of the test results on the LFW processed face dataset with or without mask occlusion and Meglass dataset are 82.50% and 89.75% respectively, which is superior to the existing algorithm, and verifies the effectiveness of the proposed method.