MFDH-Net: defect detection network for multi-level feature fusion and cross-sensing decoupling head.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Manufacturing Technology

Background:

• Industrial defect detection is crucial for quality management in intelligent manufacturing.
• Manual inspection is inefficient and prone to errors, while automated systems face challenges like inter-class similarity and multi-scale defects.

Purpose of the Study:

• To introduce MFDH-Net, an innovative network for robust industrial defect detection.
• To address challenges in detecting defects with high similarity, weak semantic information, and varying scales.

Main Methods:

• Developed a Dual-domain Feature Extraction Network (DFE-Net) for bidirectional global and local defect information capture.
• Designed a Multilevel Feature Aggregation Network (MFA-Net) with full-level fusion and adaptive weighting for deep feature interaction.
• Incorporated a Spatial Semantic Fusion Module (SSFM) for pixel-level alignment and a Cross-aware Decoupling Head (CDH) with Cross-aware Attention Module (CAM) for fine-grained recognition, especially for small defects.

Main Results:

• Achieved high mean Average Precision (mAP@.5) scores: 94.3% on steel, 96.6% on PCB, 71.7% on GC10-Net, and 98.9% on Automobile Surface Defect datasets.
• Reached a processing speed of 52 frames per second (FPS).
• Demonstrated state-of-the-art (SOTA) performance, outperforming existing methods.

Conclusions:

• MFDH-Net effectively overcomes challenges in industrial defect detection, including complex backgrounds and small targets.
• The proposed network provides a reliable technical guarantee for quality control in intelligent manufacturing.
• The results highlight the network's efficiency and accuracy for diverse industrial applications.