In modern industrial processes, a high flexibility and adaptability are indispensable. Hence, the requirements on human supporting production cells are getting even bigger.
They not only have to support the worker in the best possible manner, they also need to be easily and fast adaptable to new processes or changing conditions. As an important part of these systems, object detection and recognition provide useful information to ensure a smooth process.
In general, high accurate object detection algorithms are trained on huge datasets, like ImageNet, which takes a lot of time and computational power. In order to reduce training time during the initial setup for an industrial process, we are using a pre-trained Deep Convolutional Neural Network as a feature extractor and a fast trainable linear support vector machine for classification. The goal is to use the overhead of training in the pre-trained model to extract generalized features from a single image of the object. With the extracted features as well as negative data the classifier is trained on the fly while still obtaining an acceptable classification result.