Research on Object Detection of PCB Assembly Scene Based on Effective Receptive Field Anchor Allocation.

Area of Science:

• Computer Vision
• Machine Learning
• Manufacturing Automation

Background:

• Object detection in printed circuit board (PCB) assembly is crucial for intelligent manufacturing.
• Existing object detection models like YOLOv3 face limitations due to fixed anchor distributions that do not align with varying effective receptive fields (ERFs).
• A scarcity of dedicated datasets hinders research and development in PCB assembly object detection.

Purpose of the Study:

• To develop an improved object detection method for PCB assembly scenes.
• To address the limitations of uniform anchor distribution in YOLOv3 by incorporating ERF analysis.
• To enhance the detection of small and challenging objects like through-holes (THs).

Main Methods:

• Construction of a novel PCB assembly scene object detection dataset with 21 classes across three assembly stages.
• Refined ERF analysis on YOLOv3's feature layers to establish an ERF-based anchor allocation rule.
• Integration of an improved Atrous Spatial Pyramid Pooling (ASPP) and channel attention module to enhance detection of small objects.

Main Results:

• The ERF-based anchor allocation strategy improved mean average precision (mAP) from 79.32% to 89.86% within the YOLOv3 framework.
• The proposed method demonstrated superior performance compared to Faster R-CNN, SSD, and YOLOv4 in terms of accuracy and computational efficiency.
• Enhanced detection capabilities for small and difficult-to-detect components like through-holes were achieved.

Conclusions:

• Anchor allocation based on effective receptive field (ERF) is a viable strategy to enhance object detection accuracy in PCB assembly.
• The developed dataset and improved YOLOv3 model offer a significant advancement for intelligent PCB manufacturing.
• The proposed approach provides a balanced solution for high detection accuracy and low computational complexity in industrial applications.