Vol 15, No 1 (2024)
Recognition of cadastral coordinates using convolutional recurrent neural networks
Abstract
The article examines the use of convolutional recurrent neural networks (CRNN) for recognizing images of cadastral coordinates of objects on scanned documents of the «Roskadastr» PLC. The combined CRNN architecture, combining convolutional neural networks (CNN) and recurrent neural networks (RNN), allows you to take advantage of each of them for image processing and recognition of continuous digital sequences contained in them. During experimental studies, images consisting of a given number of digits were generated, and a CRNN model was built and studied. The formation of images of digital sequences consisted of preprocessing and concatenation of images of the digits forming them from one's own data set. Analysis of the values of the loss function and Accuracy, Character Error Rate (CER), and Word Error Rate (WER) metrics showed that the use of the proposed CRNN model makes it possible to achieve high accuracy in recognizing cadastral coordinates in their scanned images.
Program Systems: Theory and Applications. 2024;15(1):3-30
3-30
Optimal distribution of radiator area in immersion cooling systems of high-performance computing systems
Abstract
The problem of minimizing the processor temperature for a given heat flow is considered. Control is the distribution of the radiator area in contact with the coolant. This problem is equivalent to the problem of minimizing the average (over coordinate) entropy production. The distribution of the thermal load and the total radiator area are the conditions of the problem. It is shown that the optimal solution ensures the minimum processor temperature in immersion liquid cooling systems of high-performance computer.
Program Systems: Theory and Applications. 2024;15(1):31-40
31-40
Fractal model of macrosystems
Abstract
A mathematical model of a macrosystem of arbitrary nature in the form of a fractal graph is considered. This representation allows one to obtain phenomenological dependencies of macrosystems without being based on the properties of elementary objects that form the macrosystem. It is shown that a metric can be introduced on a set of stationary processes; the entropy production in the macrosystem has metric properties.
Program Systems: Theory and Applications. 2024;15(1):41-62
41-62
Justification of methods for accelerating iterative loops nests
Abstract
The acceleration of iterative algorithms, found in solving problems of mathematical physics, mathematical modeling, and image processing, is considered. In the software implementation of these algorithms, there are nested loops (sections of the program that consist of nested loops). These loop nests can be accelerated by combination of optimizing transformations, including tiling, hyperplane method, and parallelization on shared memory. The equivalence of this combination of program transformations is substantiated. A method for changing the order of tile traversal is proposed and justified. The method provides acceleration by increasing data readings from registers instead of slower memory. Considering this method, a formula for calculating optimal tile sizes is obtained. The combination of transformations presented in this article results in an acceleration that is 1.4 times greater than the well-known optimization algorithm implemented in PLUTO. In some cases using an 8-core processor, numerical experiments show a significant increase in speed compared to the original sequential algorithms. The findings of this article can be applied to manual and automatic program optimization.
Program Systems: Theory and Applications. 2024;15(1):63-94
63-94