A dynamic inference network (DI-Net) for online fabric defect detection in smart manufacturing

Abstract

Online fabric defect detection plays a critical role in the quality management of textile production. However, the high-impact and low-probability characteristics of defective samples lead to redundant design of network and hinder its real-time performance. To improve the time efficiency, this paper proposes a dynamic inference network (DI-Net) which can dynamically allocate computation resources as the complexity of image. Firstly, “AND” Gates are incorporated into the backbone to control activation of network’s function modules, allowing for dynamic adjustment of network depth. Additionally, the dynamic inference module which contains several exits with inference unit is proposed to collaborate with “AND” Gates. When sample’s confidence at specific exit satisfies the early-exit policy, the inference unit will allow it to early-exit from network and output a negative value to corresponding “AND” Gate. As a result, the output of “AND” Gate will also be negative and subsequent network will not be activated. Finally, the two-stage training strategy and exit-weighted loss function are proposed to avoid crosstalk and facilitate different exits to focus on adequate samples, enabling the efficient training of DI-Net. The experiments on the fabric dataset demonstrate that the proposed DI-Net can achieve detection precision and recall over 99% for normal samples, and approximately 95% for defective samples. Besides, its detection speed has been improved by 20%, reaching 30.1 frames per second and 20.96 m/min. This indicates that the proposed DI-Net can meet the requirements of online fabric defect detection.