METHODS OF COMPLEX OBJECTS AUTOMATIC RECOGNITION BY FORM

Abstract

The analysis of literary sources shows that the need to recognize both large and small objects is an important direction in the development of robots’ technical vision modern systems and other automation technical means, which must be able to recognize any object that has fallen into their field of vision by the measured values of attributes, assign it to a certain class, make a decision, and issue a command to the robot's manipulators. The article considers approaches to recognition with the subsequent classification of small objects by such features as the coefficient of ellipticity and the form complexity coefficient. By identifying features and their combinations for identifying similar objects, you can train a machine learning model to recognize the necessary types of patterns. A comparison of the measuring methods the areas of the projections of objects whose shape is close to rectangular, objects of a round shape, and objects that represent a long, elongated figure of a complex shape is carried out. The accuracy of measurement of area and perimeters of complex figures is estimated, errors of selected values of radii of small objects are determined depending on quantization step. The possibility of recognizing objects using traditional image processing methods or modern deep learning networks is considered: an open library for working with algorithms for computer vision, machine learning and image processing OpenCV, the latest recognition models SSD, R-FCN, Faster R-CNN, Mask R-CNN and YOLO, in the architecture of which you can see many improvements and advances in methodologies detection of objects. The advantages of using the popular Faster R-CNN recognition model, which is a combination of RPN and Fast R-CNN models, for fast recognition of small objects of complex shape are shown. It is concluded that automatic recognition systems that work according to this technique, allow to explore a variety of objects, have a sufficiently high speed, but due to the complexity of their use in real time is justified only in cases when these objects have a complex form and cannot be recognized and classified by common simpler methods and means.