Ceramic gasket defect detection based on semi-supervised domain adaptive YOLOv7_OBB
Обнаружение дефектов керамических прокладок на основе полуобучаемого YOLOv7_OBB с адаптацией к предметной области
2025-05-13
SCID: 54.1/e6nyb8jh
Discuss with AI
Kullback-Leibler divergence lossYOLOv7_OBBceramic gasket defect detectionoriented bounding boxessemi-supervised domain adaptation
Figures from the paper
Abstract (AI)
Abstract In modern manufacturing industries, defect detection of ceramic gaskets is crucial for ensuring the reliability of welding processes and the quality of products. Traditional rotating bounding box based object detection methods exhibit excessive sensitivity to boundary issues, resulting in unreliable detection accuracy in such cases. Additionally, conventional machine vision-based ceramic gasket defect detection methods typically rely on large labeled datasets for training, a process that not only consumes substantial human resources but also faces challenges of inefficiency and limited accuracy, making it difficult to meet the increasingly stringent demands of industrial production. A further challenge arises from the domain gap between training (source) data and deployment (target) environments, characterized by significant differences in image resolution, texture, background complexity, and defect morphology. These domain shifts severely hinder the generalization capability of detection models. Compounding this issue is the scarcity or complete absence of labeled samples in the target domain, which makes it infeasible to apply fully supervised learning directly. These challenges underscore the necessity of incorporating domain adaptation techniques to align the feature distributions between domains and improve detection performance in real-world settings. To address these challenges, this paper proposes a novel ceramic gasket defect detection algorithm—SSDA-YOLOv7_OBB (Semi-supervised Domain Adaptive YOLOv7_Oriented Bounding Box). This algorithm is an enhancement of YOLOv7, combining Circular Smooth Labels and Kullback-Leibler Divergence loss to optimize traditional rotating bounding boxes, while incorporating the semi-supervised domain adaptation strategy of SSDA-YOLO. Experimental results on a self-constructed dataset demonstrate that the proposed method achieves an accuracy of 78.6% in horizontal comparisons with mainstream algorithms. Under different annotation ratios of 25%, 50%, and 100%, the method consistently exhibits outstanding performance and strong practical applicability, with results of 79.7%, 78.3%, and 78.1%, respectively. These findings highlight the algorithm’s significant value for ceramic gasket defect detection and its promising adaptability to other industrial defect detection tasks.
Key Findings
1
Circular Smooth Labels and Kullback–Leibler Divergence loss are integrated to reduce boundary sensitivity and improve optimization of rotating bounding boxes.
2
Experiments on a self-constructed dataset report 78.6% accuracy in comparisons with mainstream detection algorithms.
3
The abstract indicates that the proposed approach is designed to reduce reliance on large labeled datasets while improving deployment-domain generalization.
4
The method addresses source–target domain shifts in image resolution, texture, background complexity, and defect morphology with few or no labeled target samples.
5
The paper introduces SSDA-YOLOv7_OBB for ceramic gasket defect detection, combining semi-supervised domain adaptation with oriented bounding-box detection.
Research Object
ceramic gaskets in manufacturing inspection images
Research Subject
defect detection accuracy and cross-domain generalization under variations in image resolution, texture, background complexity, defect morphology, and limited target-domain labels
Publication Details
Publication Date
2025-05-13
Journal
Publisher
ISSN
Cited by
1
Access Type
Author Information
Download PDF
Subscribe to digest