[Paper Review] A PCB Dataset for Defects Detection and Classification

Weibo Huang, Wei Peng|arXiv (Cornell University)|Jan 24, 2019

Industrial Vision Systems and Defect Detection20 references73 citations

TL;DR

The paper releases a synthesized 1386-image PCB defect dataset with 6 defect types, and presents a reference-based detection pipeline plus a CNN classifier achieving high accuracy.

ABSTRACT

To coupe with the difficulties in the process of inspection and classification of defects in Printed Circuit Board (PCB), other researchers have proposed many methods. However, few of them published their dataset before, which hindered the introduction and comparison of new methods. In this paper, we published a synthesized PCB dataset containing 1386 images with 6 kinds of defects for the use of detection, classification and registration tasks. Besides, we proposed a reference based method to inspect and trained an end-to-end convolutional neural network to classify the defects. Unlike conventional approaches that require pixel-by-pixel processing, our method firstly locate the defects and then classify them by neural networks, which shows superior performance on our dataset.

Motivation & Objective

• Create a public colorized synthesized PCB dataset with defects for detection, classification, and registration tasks.
• Provide bounding boxes and rotation information to enable localization and registration research.
• Propose a reference-based inspection method and an end-to-end CNN classifier for defect classification.
• Demonstrate dataset utility with experiments on defect detection accuracy and defect classification precision.

Proposed method

• Synthesize 1386 naked PCB images with 6 defect types: missing hole, mouse bite, open circuit, short, spur, spurious copper.
• Provide bounding boxes and rotation/orientation information for each defect.
• Use a reference comparison preprocessing pipeline (registration with SURF features, adaptive binarization, XOR localization, and morphological filtering) for defect localization.
• Train a dense-like CNN classifier (two dense blocks inspired by DenseNet) to classify defect regions within bounding boxes.
• Resize defect crops to 64x64 and perform data augmentation by shifting bounding boxes during training.
• Evaluate detection performance (P_d) and classification performance (P_c, AP_c); report timing metrics.

Experimental results

Research questions

• RQ1 Can a publicly available synthesized PCB defect dataset enable fair comparisons for detection, localization, and classification methods?
• RQ2 How effective is a reference-based AOI approach combined with CNN-based defect classification on this dataset?
• RQ3 What are the detection and classification performance metrics across the six defect types?
• RQ4 What is the computational time required for the end-to-end PCB defect inspection pipeline?

Key findings

Defect Type	Actual	Detected	P_d (%)
Missing hole	497	497	0.0%
Mouse bite	492	493	0.2%
Open circuit	482	483	0.2%
Short	491	491	0.0%
Spur	488	488	0.0%
Spurious copper	503	503	0.0%

• Defect detection achieved near-perfect results for most defect types, with P_d = 0% for missing hole, 0.2% for mouse bite and open circuit, and 0% for short, spur, and spurious copper.
• Defect classification achieved high precision, with test data P_c values: Missing hole 98.96%, Mouse bite 97.94%, Open circuit 97.74%, Short 99.48%, Spur 93.65%, Spurious copper 98.52% (Average 97.74%).
• On all samples, average precision (AP_c) reached 99.40% for the full dataset and 97.74% on test data after removing duplicates.
• The end-to-end CNN with dense connections achieved 0.989 seconds total processing time per PCB on a standard CPU/GPU setup, with registration being the most time-consuming step.