Algorithmic synthesis of computational schemes for optimization of identification and image recognition of micro-objects
I.I. Jumanov, Dilshat Djumanov, Safarov Rustam Abdullayevich · 2023
The proposed methodology for optimizing the identification of micro-objects is implemented on the basis of dynamic models, neural networks of various topologies, mechanisms for extracting redundant information structures, as well as the use of statistical, dynamic, and specific characteristics of images. A software package has been implemented for identifying, recognizing, and classifying images of micro-objects, functional modules for contour segmentation, selection of reference points, reduction of redundant fragments or features, and setting variables. The software package includes mechanisms that take into account the statistical relationships of points on the image contour, the dynamics of changes in the contour curve, the formation of point coordinate matrices, the deformation of a sequence of points in segments, and the selection of stationary sections. A comparative analysis of the effectiveness of algorithms for pre-processing images, recognition, and classification was carried out using the example of pictures of medical diagnostics, and pollen grains presented to the study in solving problems of selection and seed production of wheat grains. The software modules of neural networks are based on five types of learning algorithms, which are performed according to the supervised and unsupervised methods, the training sample is modified by the methods of vector quantization, clustering, segmentation, and the formation of a “sliding window”. The efficiency of identification of images of micro-objects was studied in the presence of “noise”.