Efficient and Reliable Surface Defect Detection in Industrial Products Using Morphology-Based Techniques

Bayraktar E.

12th International Symposium on Intelligent Manufacturing and Service Systems, IMSS 2023, İstanbul, Turkey, 26 - 28 May 2023, pp.81-94 identifier

  • Publication Type: Conference Paper / Full Text
  • Doi Number: 10.1007/978-981-99-6062-0_9
  • City: İstanbul
  • Country: Turkey
  • Page Numbers: pp.81-94
  • Keywords: Fault Detection, Image Processing, Morphology, Quality Inspection, Surface defect
  • Yıldız Technical University Affiliated: Yes


Quality is a measurement-based criteria that specifies the conformity of final products to certain rules and agreements. Monitoring product quality has always been critical, cost-effective and time-intensive process during manufacturing. Surface defects have major negative impacts on the quality of industrial products. Human inspection for visual quality control is challenging and less reliable due to the influence of physical and psychological factors on the auditor, including fatigue, stress, anxiety, working hours, and environmental conditions. Considering these risks, it is impossible for humans to deliver satisfactory stable performance over a long period of time and without interruption. Moreover, with the advancements in hardware and software, quality control can be done fast, reliable, efficient, and repeatable regardless of duration. Herein, we propose morphology-based image processing approach that enables detecting the scratch or alike defects with a width of 70 µm, which corresponds to 85 pixels of a 5181 × 5981 pixels image or 154 µm corresponding 225 pixels of a 7484 × 7872 pixels image. A scanner-based device is employed to capture the images and we combined dilation, closing, median filtering as well as gradient taking and edge detection, then contour finding. Our algorithm exceeds the performance of handcrafted feature-based methods on detecting tiny defects within very large images, which even outperforms the modern deep learning-based methods when there is not/enough training data thanks to not requiring training data. The inference time of our approach for an image is less than 1 s and consequently is capable of being exploited online surface defect detection applications robustly.