Workflow Detection with Improved Phase Discriminability

Abstract

Workflow detection is a challenge issue in the process of Industry 4.0, which plays a crucial role in intelligent production. However, it faces the problem of inaccurate phase classification and unclear boundary positioning, which are not well resolved in previous works. To solve them, this paper develops a temporal-aware workflow detection framework (TransGAN) which takes advantage of the complementarity between Transformer and graph attention network to improve phase discriminability. Specifically, temporal self-attention is firstly designed to learn the relationship between different positions of feature sequence. Then, multi-scale Transformer is introduced to encode pyramid features, which fuses multiple context cues for discriminative feature representation. At last, contextual and surrounding relations are learned in graph attention network for refined phase classification and boundary localization. Comprehensive experiments are performed to verify the effectiveness of our method. Compared to the advanced AFSD, the accuracy is improved by 2.3 % and 2.1 % when tIoU=0.5 on POTFD and THUMOS-14 dataset, respectively. Empirical study of running speed indicates that the proposed TransGAN can be deployed to real-world industrial environment for workflow detection.

Keywords

• intelligent manufacturing
• workflow detection
• self-attention mechanism
• graph relation reasoning
• transformer