研究目的
To develop advanced methods for the optical quality assurance of silicon microstrip sensors, including automated defect detection and classification using machine vision and neural network techniques.
研究成果
The developed setup and methods achieve high defect detection rates (87% with machine vision, improved to 96% with neural networks). The techniques are effective for automated optical quality assurance of silicon sensors and can be adapted to other microscopic structure inspections. Future work includes correlation with electrical tests and further optimization.
研究不足
The setup requires a cleanroom environment (ISO 4) to avoid dust interference. The camera speed is low (about one frame per second), limiting scan speed. The neural network training is computationally intensive and depends on GPU capabilities (Nvidia GTX 745 with 2000 MB RAM). Some defects may require manual verification, and the system is configuration-dependent for machine vision algorithms.
1：Experimental Design and Method Selection:
The setup includes a custom-built optical inspection system with motorized XYZ, zoom, and focus stages. Pattern recognition algorithms and neural networks are employed for defect detection and classification.
2：Sample Selection and Data Sources:
Silicon microstrip sensors from various generations, sizes, and metalization types are used, with over 1050 sensors analyzed.
3：List of Experimental Equipment and Materials:
Components include a video microscope, motorized stages (Faulhaber/Movtec SMC-300 servomotors), vacuum chuck, vacuum pump (Becker 150 mbar), light source (Starlight Roma LED3), camera (Motic 5 megapixel), and software (LabVIEW 2013, NI Vision, Darknet framework).
4：Experimental Procedures and Operational Workflow:
Sensors are placed on a vacuum chuck, aligned using fiducial marks, scanned in a snake-like pattern, and images are processed offline. Defects are detected using machine vision algorithms and neural networks.
5：Data Analysis Methods:
Image analysis includes geometric and color transformations, filtering, pattern matching, and neural network-based classification. Performance metrics such as precision, recall, and intersection over union are used.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容-
optical tube
motorized 12× zoom and motorized 3 mm fine focus tube
Navitar
Allows zoom and focus adjustments for the optical system.
-
GPU
GTX 745
Nvidia
Used for training and running neural networks.
-
servomotors
SMC-300
Faulhaber/Movtec
Provide movement for the XY-linear stage with high precision.
-
vacuum pump
150 mbar
Becker
Supplies vacuum to the chuck for holding sensors in place.
-
light source
Roma LED3
Starlight
Provides illumination for the optical inspection.
-
camera
5 megapixel microscope camera
Motic
Captures microscopic images of the sensors.
-
software
LabVIEW 2013
National Instruments
Controls the setup and data acquisition.
-
software
NI Vision
National Instruments
Provides machine vision algorithms for image analysis.
-
framework
Darknet
Implements neural network models for defect detection.
-
登录查看剩余7件设备及参数对照表
查看全部
