Fine-grained feature extraction technology innovation in the process of facial AU to expression mapping based on image processing

In recent years, the recognition of facial action unit (AU) provides a more detailed framework for expression analysis. As the smallest unit of facial muscle activity, the combination and intensity of AU can be directly mapped to specific emotions. However, the existing mapping method from AU to expression has some shortcomings in fine-grained feature extraction and poor adaptability to complex scenes. In order to solve the above problems, this study proposes an innovative method of fine-grained feature extraction based on image processing, aiming at improving the accuracy and robustness of AU-to-expression mapping. In this method, a multi-scale spatio-temporal feature fusion network (MSTF-Net) is designed, which includes three innovative modules: one is a multi-scale AU local feature extraction module, which enhances the perception ability of AU area details through the improved pyramid convolution and channel attention mechanism, The second is the AU spatio-temporal evolution modeling module, which combines 3D convolution and bidirectional LSTM network to accurately capture the spatio-temporal evolution process of AU. Thirdly, the adaptive robustness enhancement module of complex scenes introduces the mechanism of confrontation training and feature decoupling to improve the stability of the model in complex scenes such as illumination change, attitude deviation and occlusion. Experiments are carried out on two mainstream data sets, BP4D+ and DISFA. The results show that this method is superior to the existing mainstream models in the tasks of AU detection and expression recognition. On the BP4D+ data set, the average F1 score reached 0.77, which was significantly higher than the contrast method. On DISFA data set, the average accuracy rate of expression recognition reaches 83.1%, which shows stronger emotion discrimination ability. In addition, the ablation experiment verifies the contribution of each module to the overall performance, and the robustness test further proves the adaptability of the model in complex scenes. The innovative method of fine-grained feature extraction technology proposed in this study provides a new idea for the field of facial expression recognition, which has important theoretical and application value.