IMAGE PROCESSING AND DEEP LEARNING
Presented by
[Link]
[Link]
CONTENTS
➢ Introduction
➢ Manual Defect Detection Methods
➢ Image Processing based Defect Detection
➢ Machine Learning based Defect Detection
➢ Deep Learning based Defect Detection
➢Performance Comparision among existing CNN Methods
➢ Research Challenges
➢ Conclusion
INTRODUCTION
The Printed Circuit Boards (PCB’s) are vital units in electronic industry present in
almost every kind of electronic devices which made them very essential for the
current electronic world.
Nevertheless, the tiny dimension of PCB’s made the manufacturing process
complicated which resulted in crucial defect detection. In order to overcome the
complication , innumerable methods were proposed by profuse researchers
based on Image Processing, Machine Learning and Deep Learning.
Fig : Defective PCB patterns.
Fig: defective solder joints in PCBA.
MANUAL DEFECT DETECTION METHODS
At first, basic electrical tests along with defect detections are
commonly achieved by Manual Vision by experienced labourers
which resulted in high investment by the factories.
Nonetheless, it is still inefficient due to the tiny dimension which made
it complicated in detecting the defects through manual vision
Therefore, to overcome the drawback Automated Optical
Inspection(AOI) technique was introduced.
IMAGE PROCESSING BASED DEFECT DETECTION
METHODS
• To promote detection accuracy and speed for better quality at low cost,
machine vision technique or automated optical inspection technique was
introduced where reference and inspected image comparison undergoes,
obtaining the results using subtraction methods by various propositions.
• Techniques focusing on surface as well as components and solder joint defect
detection is classified into two methods:
➢Image Subtraction Method
➢Background Subtraction Method
Image Processing methods also introduced NCC as well as INCC to the
electronical industries where INCC is considered to be better in comparison.
MACHINE LEARNING BASED DEFECT DETECTION
➢Machine learning offer a new method for non-destructive quality assurance in
additive manufacturing processes.
➢Machine Learning methods like convolutional neural networks enable
computer aided defect detection in additive manufacturing.
DEEP LEARNING BASED DEFECT DETECTION
Deep Learning is the last PCB defect detection strategy to detect missing or misaligned components
using the recently developed Convolutional neural Network(CNN).
CNN automatically extract and simplify the image pre processing process to obtain accuracy and
speed.
METHODS FOR SURFACE DEFECTS:
➢Ding proposed a tiny defect detection network(TDD-net) for PCB’s based on R-CNN.
➢Adibhatla used tiny YOLOv2 network model to get faster interference speed with an accuracy of
98.2%.
➢HU and WANG proposed a method based on faster R-CNN with the backbone of residual neural
network.
➢Data augumentation method achieved a reduction in computation cost and achieved a detection of
94.2% within 0.078 seconds per image.
➢Liao replaced the darknet of higher memory consumption with the light weight network to speed
up the detection with 416X416 pixels.
➢Adibhatla considered YOLOV5 for detection because of its good structure, high speed and smaller
size than previous YOLO versions with 400X400 pixels of size.
METHODS FOR SOLDER JOINT AND
COMPONENT DEFECTS
➢YOLOv3 was considered in order to localize the solder joints.
➢Sezer and Altan combined deep learning method with image
matching to detect solder joint defects.
➢Later on LD-PCB was used which achieved the detection accuracy
of 85.8% within 0.036sec per image but was outpowered by CR-
PCB whose detection accuracy was 96.7% within 0.018 seconds.
RESEARCH CHALLENGES
TRENDS IN PCB DETECTION:
PCB’s defect detection methods are classified into 4 strategies:
➢Manual visual Inspection.
➢Industrialized AOI machine.
➢Traditional image processing.
➢Machine Learning algorithms.
The trend of using CNNs to detect defects will continue for a long time due to the rapid
development of CNN models and the computation power.
PERFORMANCE COMPARISION AMONG EXISTING CNN MODELS
MODELS PARAMETERS WEIGHT mAP INFERENCE SPEED
(MB) (frames per
second)
Faster RCNN- 41.15 315 90.1% 2
ResNet50-FPN
Faster RCNN- 60.14 460 90.6% 1.4
ResNet101-FPN
RetinaNet- 36.21 277 94.4% 1.9
ResNet50-FPN
RetinaNet- 55.2 422 96.2% 1.4
ResNet101-FPN
SSD 24.41 186 86.7% 1.4
SSD-Lite 3.1 24.1 86.0% 9.8
YOLOv3-tiny 8.68 16.6 99.4% 19
YOLOv5-small 7.03 13.8 99.4% 14
CURRENT CHALLENGES FOR DEEP LEARNING
METHODS:
➢ OPEN PCB DEFECT DATASET
The existing open datasets mainly focus on component detection or cosmetic defect
detection with only one type of defect in one image.
➢TINY COMPONENT DEFECT DETECTION
With the development of microelectronic technology, the assembled PCB is becoming
increasingly smaller. Hundreds of tiny components can be mounted on a small board,
which is difficult for human eyes to inspect.
➢SOLDER JOINT DEFECT DETECTION
Without assistance, they can only detect the qualified and defective classification with
good performance due to which the detecting classes are limited to a small range, not
coinciding with the real industrial manufacturing process.
CONCLUSION:
Hereby, we conclude that with multi-level feature extraction
and automatic learning abilities, CNN- based defect
detection algorithms have overcome such issues and
achieved significant performances in both accuracy and
speed.
ANY QUESTIONS