Facial expressions are a vital way for humans to show their perceived emotions. It is convenient for detecting and recognizing expressions or micro-expressions by annotating a lot of data in deep learning. However, the study of video-based expressions or micro-expressions requires that coders have professional knowledge and be familiar with action unit (AU) coding, leading to considerable difficulties. This paper aims to alleviate this situation. We deconstruct facial muscle movements from the motor cortex and systematically sort out the relationship among facial muscles, AU, and emotion to make more people understand coding from the basic principles: We derived the relationship between AU and emotion based on a data-driven analysis of 5,000 images from the RAF-AU database, along with the experience of professional coders. We discussed the complex facial motor cortical network system that generates facial movement properties, detailing the facial nucleus and the motor system associated with facial expressions. The supporting physiological theory for AU labeling of emotions is obtained by adding facial muscle movements patterns. We present the detailed process of emotion labeling and the detection and recognition of AU. Based on the above research, the video's coding of spontaneous expressions and micro-expressions is concluded and prospected.

Image anomaly detection and localization have received widespread interest in the community, and knowledge distillation (KD) has been widely explored. Recently, the reverse distillation (RD) paradigm has successfully mitigated the homogenization of anomaly representations in traditional KD arising from identical or similar teacher-student (T-S) architecture. However, in RD, the lack of an effective means to prevent anomalous patterns in the teacher encoder from being leaked into the student decoder undermines potential modeling discrepancies between the T-S model in anomaly representations. To settle this problem, we propose REverse distillation with latent Anomaly SuppressiON (REASON) method, preventing the student decoder from receiving anomalous patterns by extra means of anomaly filtering during the inference phase, and thus, the student model can only restore representations of anomaly-free images. Specifically, we construct a Siamese teacher encoder architecture, with one branch extracting features from anomaly-free samples and the other synthesizing anomaly features with spatial noise injection from the latent feature level. Next, we design a latent anomaly suppression module to recover normal features from perturbed anomalous features. In this sense, the follow-up student decoder will receive input without abnormal patterns. Thus, representations of the anomaly-free images can be described well, while those of the anomalous images can be well-differentiated between the T-S model. Furthermore, to enhance the model’s anomaly detection and localization capabilities, we propose multi-granularity KD loss to optimize the student decoder to focus on context and local details. Extensive experiments are performed on three benchmark datasets, i.e., MVTec AD, AeBAD, and OCT2017, and the results show the effectiveness and robustness of our proposed approach, which achieves superiority over the current state-of-the-art methods.

Micro-expression recognition (MER) is inherently challenging due to the difficulty of extracting subtle, localized changes in micro-expressions (MEs). Various optical flow-based methods have been proposed for MER, as optical flow can effectively suppress facial identity information while capturing the movement patterns of MEs. However, these methods, characterized by simple architectures, often fail to extract discriminative features, resulting in suboptimal performance. In this paper, we propose an Enhanced Micro-expression Recognition Network with attention and distance correlation (EMRNet) for MER. EMRNet consists of three key phases: First, we introduce a novel channel-wise region-aware attention mechanism within two identical Inception networks, designed to extract global and local expression features in parallel, based on the optical flow input of the same ME. Second, to enhance ME representations, we propose a regularized dilated loss function incorporating distance correlation, which improves the information entropy transferred between the two branches. Last, emotion categories are predicted by fusing the expression-dilated features in the classification branch. Extensive experiments conducted on the composite database from the MEGC 2019 challenge demonstrate the effectiveness of EMRNet under both leave-one-subject-out (LOSO) cross-validation and the composite database evaluation (CDE) protocol. The results show that our approach successfully generates discriminative features, achieving substantial performance gains. Furthermore, EMRNet outperforms existing single-stream and dual-stream models, delivering superior results in MER.