Defect Detection Application

The project is assigned by FAREL Plastic that is a joint company partnered by Farplas and Wirthwein DE.

This project aims to detect some specific surface defects on plastic surfaces. The final product targets operation on a conveyor belt. This code includes MATLAB modeling (with deep learning), final algorithm (conventional computer vision) and GUI (.NET) in C++ programming language.

The problem is clearly stated below, the defects are hard to be spotted by naked eye. So, it is best practice to adjust photometric settings of the lens & camera to acquire some data containing defect information.

Fig. 1: Two kinds of surface defects that are slightly whiter areas (e.g. the one on the left can be depicted as semi-ellipse)

Hardware

• Lens: Fujinon HF8XA-5M
• Camera: TIS DMK 27AUP031 (QE: best in 400-600 nm range)
• Sensor: CMOS Aptina MT9P031
• Connection: USB 3.0 SuperSpeed

Dependencies

• Language standard - C++14
• .NET Framework(C++/CLI) v4.7.2
• OpenCV v4.5.0
• Camera library:
  • tiscamera (for Linux)
  • TIS_UDSHL12_x64 (for Win, packed in SDK)
• Camera driver: TIS Camera v2.9.8
• Camera C++ SDK: IC C++ v3.5.7

Warning In case of broken URLS, all the software are likely to be found here.

Installation

Simply download content in /bin/Release/ and follow the steps from setup.exe. Only Windows x64 installation is provided. The camera driver has to be installed seperately. While, other dependencies are development related.

MATLAB Experiment (learning-based)

This block under /src/matlab/ section implements an Faster-RCNN based training to automatically detect and locate specific defect types. Although not used by this project, a MATLAB toolbox for TIS Camera can found here.

matlab_result.mp4

Fig. 2: Learning-based experiment (white pixels are candidate features, bounding box gives the most probable defect)

GUI (filter-based)

GUI is designed and compiled in .NET Framework. The algorithm leveraged conventional filter-based computer vision techniques, powered by OpenCV. In contrast to automated bounding boxes of learning-based simulation, this implementation needs user input for fixed ROIs. It is due to the fact that the system is assumed to stable on a conveyer belt; as a result, location automation is discarded to increase performance.

gui_demo.mp4

Fig. 3: GUI demonstration by adjusting user defined ROI and calibration