This paper presents a new methodology for analyzing lip articulation during fingerspelling aimed at extracting robust visual patterns that can overcome the inherent ambiguity and variability of lip shape. The proposed approach is based on unsupervised clustering of lip movement trajectories to identify consistent articulatory patterns across different time profiles. The methodology is not limited to using a single model. Still, it includes the exploration of varying cluster configurations and an assessment of their robustness, as well as a detailed analysis of the correspondence between individual alphabet letters and specific clusters. In contrast to direct classification based on raw visual features, this approach pre-tests clustered representations using a model-based assessment of their discriminative potential. This structured approach enhances the interpretability and robustness of the extracted features, highlighting the importance of lip dynamics as an auxiliary modality in multimodal sign language recognition. The obtained results demonstrate that trajectory clustering can serve as a practical method for generating features, providing more accurate and context-sensitive gesture interpretation.

The image processing method involves several critical steps, with image preprocessing being particularly significant. Segmentation and contour extraction on digital images are essential in fields ranging from image recognition to image enhancement in various recording devices, such as photo and video cameras. This research identifies and analyzes the main drawbacks of existing segmentation and contour extraction methods, focusing on object recognition. Not all filters effectively remove noise; some may clear areas of interest, affecting gesture recognition accuracy. Therefore, studying the impact of image preprocessing on gesture recognition outcomes is crucial for improving pattern recognition performance through more efficient preprocessing methods. This study seeks to find an optimal solution by detecting specific features during the preprocessing stage that directly influence gesture recognition accuracy. This research is a key component of the AP19175452 project, funded by the ministry of science and higher education. The project aims to create automated interpretation systems for Kazakh sign language, promoting inclusivity and technological innovation in communication aids. By addressing these challenges, the study contributes to the development of more robust and adaptive image preprocessing techniques for gesture recognition systems.

The multilingualism found in many countries, as well as within professional groups, complicates verbal communication, as both communicating parties are required to know all the languages used. This problem is exacerbated by the fact that languages are often mixed during communication. Avatars can be used to communicate with deaf people by simulating the behavior of sign language users. This paper presents a digital sign language avatar built on a language-agnostic, multimodal animation pipeline that decouples linguistic input from animation, combining skeletal body and hand motion with facial blendshape animation as independent modalities. It also presents a methodology for assessing its quality with the participation of experts (i.e., professional sign language interpreters) and the corresponding research results. The average quality rating of the avatar interface by the experts was 5.5 on a 7-point Likert scale, indicating its potential for practical use. At the same time, the research identified opportunities to improve the naturalness of movement and the consistency of gesture transitions.

This article investigates the effects of dataset size and class imbalance on the performance of You Only Look Once (YOLO)v10-based personal protective equipment (PPE) detection systems. Six YOLOv10 configurations (n, s, m, b, l, and x) were evaluated using domain-specific datasets from construction, industrial, and medical contexts. A novel mathematical model is proposed to characterize the nonlinear relationship between dataset size and detection performance (mean Average Precision (mAP)@50), revealing a saturation threshold beyond which additional data yield diminishing performance gains (R² = 0.981). The analysis further demonstrates that the superior accuracy of more complex YOLOv10-x models deteriorates significantly under conditions of pronounced class imbalance. These findings underscore the critical importance of balanced and sufficiently large datasets for optimizing detection accuracy in real-world safety applications.