Data fusion strategy to improve the realiability of machine learning based classifications

Abstract

Automatic object recognition plays a major role in many industrial applications. This task is mostly performed by using optical sensors and image processing methods. Degeneration processes, such as surface wear, however, can pose quite some challenges when it comes to high-quality optical recognition. In this article we present our solution to optical character recognition of strongly degenerated numbers, characterized by a varying embossing depth and texture intensity, imprinted on metal surfaces. Under these conditions Machine Learning (ML) based recognition models seem to perform better than conventional ones. Typically, ML models have a black box character in the sense that the algorithm steps have no direct interpretable meaning and are kind of arbitrary. Consequently, the results of such models are difficult to interpret with respect to their trustworthiness. In order to receive more reliable recognition results, we have developed a rule-based fusion strategy that combines the output of several different AI models. This approach not only leads to a higher rate of correctly recognized objects, it also indicates when the recognition result is uncertain. As a result, our method increases the process safety and makes object recognition in industrial applications more flexible and robust.