Search academic texts

Information × Registration Number 0225U002621, (0122U200629) , R & D reports Title Machine learning and standard feature descriptors in pattern recognition popup.stage_title Машинне навчання та стандартні дескриптори ознак у розпізнаванні образів Head Furhala Yurii M., Кандидат фізико-математичних наук Registration Date 23-03-2025 Organization Ivan Franko National University of Lviv popup.description1 Evaluation of the effectiveness of combining characteristic methods of describing objects with classification methods based on machine learning and comparing the results with classical recognition methods. Development of criteria for their optimal combination. popup.description2  Object of research: Methods of detection, description and filtering of keypoint correspondences in images, in particular the SIFT, SURF, ORB, BRISK algorithms, in combination with the classical RANSAC method and adaptive USAC variants. Comparison of keypoint detection algorithms with the support vector method and neural networks. Study of the impact of various distortions on the color characteristics of images (rotation, distortion, blurring and noise superimposition). Purpose of the work: analysis of the effectiveness of methods of detection, description of keypoints and correspondence filtering. Study of the possibility of using color histograms formed in HS* color spaces for identification (comparison) of images with superimposed distortions. Research methods: theoretical analysis of modern algorithms, modeling of their operation using the OpenCV library and experimental evaluation of efficiency on test image sets. Correlation analysis of color histograms of original and distorted images. Results obtained. It is shown that the SIFT method provides the highest accuracy and stability, the SURF and BRISK methods achieve a balance between accuracy and speed, and the ORB method demonstrates the highest speed, which is an advantage for limited resources. The USAC_PROSAC and USAC_FAST methods have the fastest processing, while the USAC_MAGSAC method provides stability for complex tasks. The RANSAC method was distinguished by the highest stability, and optimization of its parameters improved accuracy without loss of performance. Keypoint-based methods in combination with SVC provide high speed and are appropriate for use in conditions of limited computing resources. The YOLOv5 neural network offers higher accuracy in complex visual scenarios. The results of the conducted studies suggest the admissibility of using color histograms to identify images or their fragments even with fairly high degrees of their distortion by interference of various nature. Product Description popup.authors Vdovychenko Viktor M. Velhosh Andrii S. Luchka Vasyl-Taras Ya. Fesiuk Andrii V. popup.nrat_date 2025-03-23 Close