Метод Асинхронної Конвеєрної Обробки Потокових Зображень Для Семантичної Сегментації В Реальному Часі
Year: 2026
Type: article
Abstract: An asynchronous pipelining method for streaming images using a central processing unit and a graphics processing unit is developed for real-time semantic segmentation. The method addresses the accumulation of local delays when preprocessing, data transfer, denoising, contour-oriented feature enhancement, neural-network segmentation, and output formation are executed as weakly coordinated sequential stages. The proposed organization preserves causal dependencies within each image portion and simultaneously processes different stream portions at different pipeline stages. Circular buffering, asynchronous host-to-device and device-to-host transfers, graphics processing unit streams, and readiness events are used to coordinate execution without global synchronization barriers. After pipeline filling, the stationary output interval is determined by the dominant stage rather than by the sum of all stage durations. Experimental verification showed that parallel preprocessing reduced data preparation time by 89.06 % and increased input throughput to 6443.12 images/s. The graphics processing unit wavelet-denoising stage processed more than 200 images/s while preserving reconstruction quality. Sobel-based contour-oriented enhancement increased the Dice coefficient by 1.42 percentage points for the lungs, 2.18 percentage points for the heart, and 6.37 percentage points for the clavicles. Sequential inclusion of this stage increased processing time from 2.50 to 2.87 s/image and reduced throughput from 0.400 to 0.348 images/s. In the asynchronous pipeline, the same enhancement remained in the full route while the stationary output interval stayed at 2.50 s/image because semantic segmentation remained the dominant stage. The method therefore improves segmentation quality without reducing stationary throughput relative to the baseline route. The coordinated organization also prevents non-dominant local stages from unnecessarily reducing the stable mask-generation rate while preserving causal correctness of the complete processing route.
Author: Гентош, Леся
Institution: Lviv Polytechnic National University
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214832946