Software & Systems

Articles of journal № 3 at 2020 year.

Order result by: Public date | Title | Authors |

21. Technical object projection in the image into metric space using deep neural networks for the detection problem [№3 за 2020 год]
Authors: Tolstel, O.V., Shirkin, A.E., Kalabin, A.L.
Visitors: 3871
The paper presents the image vectorization algorithm containing technical objects. As technical objects are machine-building parts, fasteners, hardware. Vectorization refers to the transformation of an image into a vector for which the Euclidean distance has semantic meaning. This algorithm was created to im-prove the system of assessing the object position, where there is a problem of a variable number of types of objects for recognition. The author proposes an approach to the formation of a metric space for images, where the image transformed into a vector by the metric l2 can be compared with the image by the standard, thereby solving the problem of a non-constant number of classes. To add a new class, it is enough to add a ref-erence image represented as a vector to the system and find the distance to it. If it is smaller than other images, then this reference will represent the type of object that was submitted to the input system. This approach is implemented in deep neural networks, where the last layer is removed and the penultimate layer is left, which represents the upper level of features extracted from the image. Such a neural net-work undergoes a learning process using the Triplet loss function, teaching the neural network to vec-torize the image into metric space. The program implementing the proposed algorithm is developed in Python 3.6 using the Jupyter Lab integrated environment for the Ubuntu 18.04 operating system. The results of the experiment on the use of the proposed algorithm, which consisted in attributing the obtained images to a particular image - standard, are presented. To assess the quality of the algo-rithm, ranking metrics were used for search problems, where only the very first object in the list of nearest objects is evaluated. The developed algorithm can be used for technical vision systems for robotic manipulators, and in the future this algorithm will be used as part of the control system for capturing objects by a robotic manipulator.

◄ ← Preview | 1 | 2 | 3