Research on Reconstruction Method of Occluded Micro-Expressions Based on Quantized Denoising Autoencoder
Micro-expressions serve as a crucial basis for psychological health diagnoses, and occlusions caused by objects, such as glasses or masks, can make micro-expression recognition challenging. Existing reconstruction methods for occluded micro-expressions rely primarily on reconstructing RGB texture in...
Uloženo v:
| Vydáno v: | Ji suan ji gong cheng Ročník 51; číslo 5; s. 288 - 304 |
|---|---|
| Hlavní autor: | |
| Médium: | Journal Article |
| Jazyk: | čínština angličtina |
| Vydáno: |
Editorial Office of Computer Engineering
15.05.2025
|
| Témata: | |
| ISSN: | 1000-3428 |
| On-line přístup: | Získat plný text |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
| Shrnutí: | Micro-expressions serve as a crucial basis for psychological health diagnoses, and occlusions caused by objects, such as glasses or masks, can make micro-expression recognition challenging. Existing reconstruction methods for occluded micro-expressions rely primarily on reconstructing RGB texture information, which leads to issues such as information redundancy and difficulties in achieving precise texture reconstruction. In addition, the models used in such reconstruction methods often involve symmetric autoencoders based on U-Net and Generative Adversarial Networks (GAN); However, the former suffers from limited reconstruction capabilities in shallow symmetric structures, and the latter faces challenges in terms of adversarial loss convergence speed. This paper proposes a method for reconstructing dynamic flow features in occluded regions of micro-expressions based on a vector-quantized denoising autoencoder. First, dynamic flow, a robust feature representation resilient to lighting variations, is proposed based on optical flow and dynamic images, effectively aggregating motion information from all TVL1 optical flows and simplifying the texture information. Then, a two-pair Vector Quantized Denoising Autoencoder (VQ-DAE) based on the discrete encoding Vector Quantized Variational Autoencoder (VQ-VAE) is introduced to reconstruct dynamic flow features in occluded regions of micro-expressions to facilitate the recognition of occluded micro-expressions. Experimental results demonstrate that this approach effectively reconstructs motion information in occluded regions, achieving accuracy rates of 77.89%, 72.02%, and 61.04% on the CASME, CAS(ME)2, and CASME Ⅱ datasets, respectively. Compared to traditional, spatial-attention-, and self-attention-based methods, our method leads to significant improvements in accuracy, Unweighted Average Recall (UAR), and Macro-F1. |
|---|---|
| ISSN: | 1000-3428 |
| DOI: | 10.19678/j.issn.1000-3428.0068949 |