Multidimensional computed measurement for highly accurate PCBA defect detection.

Area of Science:

• Industrial Automation
• Computer Vision
• Quality Control

Background:

• Automated optical inspection (AOI) is vital for industrial quality control.
• Current AOI methods often struggle with accuracy due to reliance on single-texture image data, lacking 3D topographic information.
• This limitation hinders robust defect detection in complex manufacturing environments.

Purpose of the Study:

• To develop a more accurate and robust defect detection method for industrial AOI.
• To overcome limitations of existing techniques by integrating multidimensional image information.
• To enhance the precision of defect identification in printed circuit board assembly (PCBA).

Main Methods:

• A novel multidimensional information fusion (MIF) module was proposed, integrating texture and depth image features.
• The MIF module employs specialized mechanisms for spatial and channel feature extraction and a hierarchical fusion strategy using Transformer blocks.
• A position information mask (PIM) module was introduced for post-processing, utilizing Gerber file data to filter misidentified defects.

Main Results:

• The proposed method achieved an average accuracy of 99.93% on a PCBA defect dataset.
• This represents a significant improvement of 5.64% over conventional YOLOV5 methods.
• Ablation studies confirmed the effectiveness of the MIF and PIM modules in enhancing detection performance.

Conclusions:

• The developed method significantly improves accuracy and robustness in PCBA surface defect detection.
• The integration of texture and depth information via the MIF module is key to enhanced performance.
• The PIM module effectively refines detection by incorporating design-specific positional data, offering a valuable reference for industrial AOI.