Methods for Recognizing Images of Heterogeneous Objects in Small Training Sample
Iurii E. Shishkin, Aleksandr N. Grekov · 2020 International Multi-Conference on Industrial Engineering and Modern Technologies (FarEastCon) · 2020
The article studies the application of computer vision, digital filtering and two-dimensional convolution methods for solving image recognition issues of heterogeneous objects by the example of the in situ plankton recognition in a small training sample. Key focus is on the study of the stability of computer vision template matching methods toward a group of affine transformations. The basic provisions are given alongside with the mathematical substantiation of the application of the methods of two-dimensional convolution with a dynamically generated kernel in order to solve the image recognition problem. The authors carry out a series of targeted synthetic tests for the recognition of plankton images in the conditions of unstable environment and the presence of random geometrical distortions. Moreover, for each class of the training sample, the minimum number (not more than ten) of manually prepared patterns is used. The Euclidean distance and correlation coefficient are used as classifier metrics. It is shown that the developed system successfully solves the issues of detecting objects against an unstable background and determining their class under complex conditions: multiple objects scale, random location and rotation, point and wave noise components on the background image texture, as well as minor brightness changes.