Feature Point Extraction Model for Improving Semiconductor Package Inspection Efficiency


  • Yeong-Keun Lee
  • Koo-Rack Park
  • Jae-Woong Kim
  • Dong-Hyun Kim


Background/Objectives: With the development of IT, the manual work has been changed to the mechanical work. As smart factories have been spread widely, there has been on the sharp increase in the demand of applying vision system to automation process. In particular, the semiconductor production process utilizes vision system as an image processing and image analysis technology in order for efficient control and processing. Accordingly, it is necessary to develop the vision system that makes it possible to inspect whether products pass or fail accurately and immediately.
Methods/Statistical analysis: With the high density and minimization of such products as semiconductors, the human inspection with the naked eye can cause a slow inspection speed and low efficiency and accuracy. In addition, since the inspection criteria for products are ambiguous in such an inspection way, it is impossible to guarantee product quality. This study implemented the model that makes it possible to extract feature points of objects through ORB algorithm to inspect whether products pass or fail in the BGA of semiconductor package, to classify and save image data through SVM algorithm, and thereby to improve product inspection efficiency through vision system.
Findings: IC logic used in semiconductor manufacturing is an expensive chip with high performance. Therefore, if a package defect is found in the middle of manufacturing, it causes a lot of losses. Removing defected products in the production line through vision inspection system can greatly influence improvements in the final value and reliability of products. The proposed system is the SVM algorithm based semiconductor package inspection model that makes it possible to classify and save the semiconductor package images put in by probe with the application of pre-processing process and ORB algorithm for feature point extraction. In the experiment of the proposed model, when one obtained image was learned, recognition rate was about 28%; when 14 images were learned, recognition rate was 97%. If the proposed model is applied to vision system as semiconductor inspection equipment, it is expected to inspect products more quickly and accurately.
Improvements/Applications: The proposed model makes it possible to not only inspect semiconductor package, but extract feature points. Therefore, it is possible to apply the model to text or symbol recognition. In the future, it will be necessary to research how to apply a different algorithm to extract feature points, and how to shorten the time taken to detect defected products in order to inspect products immediately and accurately.