Convolutional Neural Networks for Image Recognition in Mixed Reality Using Voice Command Labeling

摘要

In the context of the Industrial Internet of Things (IIoT), image and object recognition has become an important factor. Camera systems provide information to realize sophisticated monitoring applications, quality control solutions, or reliable prediction approaches. During the last years, the evolution of smart glasses has enabled new technical solutions as they can be seen as mobile and ubiquitous cameras. As an important, aspect in this context, the recognition of objects from images must be reliably solved to realize the previously mentioned solutions. Therefore, algorithms need to be trained with labeled input to recognize differences in input images. We simplify this labeling process using voice commands in Mixed Reality. The generated input from the mixed-reality labeling is put into a convolutional neural network. The latter is trained to classify the images with different objects. In this work, we describe the development of this mixed-reality prototype with its back-end architecture. Furthermore, we test the classification robustness with image distortion filters. We validated our approach with format parts from a blister machine provided by a pharmaceutical packaging company in Germany. Our results indicate that the proposed architecture is at least suitable for small classification problems and not sensitive to distortions.