Introducing a Decision Support System for Use-Case Specific Object Detection Methods in Production Systems

Abstract

Object detection has the potential to facilitate the automation of optical quality inspection, achieving a significant reduction of human error. However, the gap between needed expertise to understand and integrate complex object detection systems into production environments and the availability of computer scientists, is hindering its use in the manufacturing industry. A support system for decision-makers unrelated to the subject is therefore required to promote industry utilisation of object detection and effectively manage this otherwise unused opportunity of profitable knowledge. Within the Cluster of Excellency “Internet of Production (IoP)”, such a support system has been developed. Lowering the implementation hurdle of object detection systems is achieved by translating complex information about existing methods into tangible factors, such as quality and cost. In this work we aim to structure relevant object detection techniques and employ a decision tree to provide a user-support based on a use-case-oriented framework. The use-case’s basic conditions and requirements serve as input for the framework. The decision tree gives the suitable object detection method as output, accordingly. For traditional object detection techniques, the characteristics are translated into basic requirements, which are then used as input, e.g., for the template matching method the comparison of a source image with a reference image is translated to the ability to guarantee images for all possible quality deviations. The deep learning (DL) methods are consolidated in the project management triangle consisting of quality, cost, and time. Firstly, an introduction into object detection is given. Secondly, traditional methods are clustered and deep learning methods classified. A description of the decision tree is then presented, before testing results conclude this paper. The developed support system enables decision-makers to evaluate object detection methods for individual use-cases and consequently achieve increased production planning efficiency. The system’s universal design allows for application across manufacturing industries and use-cases.