Method of Automatic Images Datasets Sampling for the Manual Operations Control Systems

The paper presents a complex method for automatic construction of data samples for systems of intelligent control of manual operations in industrial production. In the described systems, objects and operations are recognized using ML-methods, primarily neural networks. Their effectiveness depends on the quality and size of training data sets. At the same time, the time spent on the gathering, processing, marking of such data sets is a significant part of the total time of the system implementation. To solve this problem, the authors proposed a method for automatic sampling. It allows to significantly reduce the time required to construct new data sets when configuring the system to control a new technological process.The main idea of the proposed method is that the operator demonstrates the assembly parts one by one in front of the camera, and the system automatically creates a marked-up set of images. The article describes the progress and the results of the research. The results of experimental comparisons of algorithms for selecting objects from the background are given. The procedures for filtering "defective" images are described. Examples of automatically created data sets are shown, as well as the results of their use for training neural networks within the framework of the problem being solved. The final tests were carried out using a specially designed software and hardware operation control demo-stand. The tests showed the high efficiency of the developed method.