Small sample pipeline DR defect detection based on smooth variational autoencoder and enhanced detection head faster RCNN

The safe operation of gas pipelines is crucial for the safety of residents’ lives and property. However, accurately detecting defects within these gas pipelines is a challenging task. To improve the accuracy of defect detection in pipeline DR images with small sample sizes, we propose an enhanced Fa...

Full description

Saved in:
Bibliographic Details
Published in:Applied intelligence (Dordrecht, Netherlands) Vol. 55; no. 10; p. 716
Main Authors: Zhang, Ting, You, Tianyang, Liu, Zhaoying, Rehman, Sadaqat Ur, Shi, Yanan, Munshi, Amr
Format: Journal Article
Language:English
Published: Boston Springer Nature B.V 01.06.2025
Subjects:
Accuracy
Adaptive sampling
Classification
Datasets
Defects
Gas pipelines
ISSN:0924-669X, 1573-7497
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The safe operation of gas pipelines is crucial for the safety of residents’ lives and property. However, accurately detecting defects within these gas pipelines is a challenging task. To improve the accuracy of defect detection in pipeline DR images with small sample sizes, we propose an enhanced Faster RCNN model based on a Smooth Variational Autoencoder and Enhanced Detection Head (S-EDH-Faster RCNN). This model leverages a smooth variational autoencoder to reconstruct features and enhances classification scores through an improved detection head, thereby boosting overall detection accuracy. In detail, to address the issue of scarce training samples for new categories, we design a smooth variational autoencoder to reconstruct features that better fit the distribution of training data. Furthermore, to refine classification precision, we present an enhanced detection head that incorporates a convolutional block attention-based center point classification calibration module, which strengthens classification-related portions of the RoI features and adjusts classification scores accordingly. Finally, to effectively learn characteristics of novel class samples, we introduce an adaptive fine-tuning method that adaptively updates key convolutional kernels during the fine-tuning stage, enabling the model to generalize better to novel classes. Experimental results demonstrate that our approach achieves superior detection performance over state-of-the-art models on both the home-made PIP-DET dataset and the publicly available NEU-DET dataset, demonstrating its effectiveness.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0924-669X
1573-7497
DOI:10.1007/s10489-025-06590-3