Liquid Reservoir Weld Defect Detection Based on Improved YOLOv8s.

Area of Science:

• Mechanical Engineering
• Computer Vision
• Materials Science

Background:

• Weld seam defects on automotive liquid reservoirs pose quality control challenges.
• Traditional detection methods are insufficient for varied defect shapes and scales.

Purpose of the Study:

• To develop an advanced algorithm for accurate detection of liquid reservoir weld defects.
• To improve the performance of the YOLOv8s model for detecting defects with size variations and complex features.

Main Methods:

• Proposed an optimized YOLOv8s algorithm incorporating Reparameterized Generalized Feature Pyramid Network (RepGFPN) and a small-object detection head.
• Replaced Spatial Pyramid Pooling Fast (SPPF) with Focal Modulation Networks (FocalNets) for complex defect identification.
• Integrated Cascaded Group Attention (CGA) mechanism to reduce redundant feature information propagation.

Main Results:

• Achieved a 6.3% improvement in mAP@0.5 and 4.3% in mAP@0.5:0.95 over the original YOLOv8s model.
• Demonstrated significant AP improvements for craters (3.9%), porosity (13.5%), undercuts (5.0%), and lack of fusion (2.5%).
• Outperformed state-of-the-art models on both liquid reservoir and steel pipe weld defect datasets.

Conclusions:

• The enhanced YOLOv8s algorithm provides superior performance for detecting diverse weld defects in automotive components.
• The proposed optimizations effectively address challenges related to defect size variation and complexity.
• This method offers a robust solution for automated quality inspection in manufacturing.