Моделі нормалізації та розпізнавання зображень

Abstract

This work is devoted to the investigation of image normalization models and recognition methods, particularly for the task of clothing type classification and logo detection on garments in real-world conditions. The main approaches to image preprocessing, including various normalization techniques, standardization methods, and data augmentation strategies, as well as modern deep learning architectures such as convolutional neural networks (CNN), residual networks (ResNet), and object detection models (YOLO series) were considered. A comparative calculation of the effectiveness of single-stage and multi-stage normalization pipelines relative to non-normalized images was carried out for datasets typical for fashion recognition tasks, such as DeepFashion, Fashion Product Images, and LogoDet-3K. It has been found that the two-stage normalization system (pixellevel scaling combined with batch normalization and advanced augmentation) is the most effective compared to other approaches, especially under conditions of variable lighting, occlusion, and small logo sizes.