A Lightweight Pipeline Edge Detection Model Based on Heterogeneous Knowledge Distillation

The pipeline safety warning system (PSEW) is an important guarantee for the safe transportation of energy pipelines. Given the constraints of deploying detection models at resource-limited pipeline stations, there is a compelling need to develop efficient, lightweight models suitable for edge device...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on circuits and systems. II, Express briefs Express briefs, 2024-12, Vol.71 (12), p.5059-5063
Main Authors: Zhu, Chengyuan, Pu, Yanyun, Lyu, Zhuoling, Wu, Aonan, Yang, Kaixiang, Yang, Qinmin
Format: Article
Language:English
Subjects:
Adaptation models
Computational modeling
distributed optical fiber
Feature extraction
heterogeneous model
Image edge detection
Integrated circuit modeling
knowledge distillation
lightweight
Optical fiber networks
Pipeline network safety
Pipelines
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The pipeline safety warning system (PSEW) is an important guarantee for the safe transportation of energy pipelines. Given the constraints of deploying detection models at resource-limited pipeline stations, there is a compelling need to develop efficient, lightweight models suitable for edge device applications. This brief introduces an adaptive heterogeneous model knowledge distillation network (AHKDnet) for edge deployment of pipeline network detection models. The global information and long-distance dependency relationships from the ViT-based teacher network are transferred to the CNN-based shallow student network. We introduce the learnable modulation parameters to optimize target information enhancement, reducing the impact of irrelevant information. By embedding the model selection at each stage of knowledge distillation, the performance collapse of student models caused by misleading cross-architecture knowledge is avoided, and model convergence is accelerated. Experiments on three actual scene datasets of pipeline networks show that AHKDnet outperforms the state-of-the-art KD methods and has strong generalization ability. Notably, AHKDnet enhances the recognition performance of shallow student networks by an average of 10%, highlighting its efficacy and potential for practical applications. Our method can provide a new reference for edge deployment of PSEW.
ISSN:1549-7747
1558-3791
DOI:10.1109/TCSII.2024.3439361