ItemArticle

High-speed portrait video segmentation using the hybrid combination of deep-learning models and boundary movement adjustment

Title
High-speed portrait video segmentation using the hybrid combination of deep-learning models and boundary movement adjustment
Creator
Kim, Yong-Woon; Byun, Yung-Cheol; Krishna, Addapalli V.N.; Krishnan, Balachandran
Description
As global warming intensifies, the development of energy-efficient Artificial Intelligence (AI) technologies has become crucial. Additionally, the growing demand for on-device AI in smartphones, extended reality devices, and autonomous vehicles necessitates AI systems that can function effectively on low-performance hardware. To address these needs, this study proposes hybrid methods in the field of Portrait Video Segmentation (PVS). Our proposed hybrid models leverage Deep-learning based Segmentation Models (DSMs) and a novel Boundary Movement Adjustment (BMA) process to achieve speed and accuracy balance. The Hybrid Serial Model (HSM) not only accelerates PVS but also improves energy efficiency while maintaining a similar level of accuracy. On the other hand, the Hybrid Parallel Model (HPM) enables high-performance PVS even on low-performance devices, ensuring no video frames are lost during high-speed segmentation processing. Tests conducted on Jetson Nano, Raspberry Pi, and a desktop PC demonstrate the effectiveness of these models, showing improvements in PVS speed while maintaining accuracy close to that of traditional DSMs. HSM increased PVS speed from 15.2 Frames Per Second (FPS) to 25.1 FPS on a desktop PC with a 0.5 % accuracy loss, and from 6.3 FPS to 16.5 FPS on a Jetson Nano with a 1 % loss. HPM reached 30 FPS on a desktop PC with a 0.05 % loss, and 29.7 FPS on a Jetson Nano with a 1 % loss. On the Raspberry Pi, the HPM method improved speed from 2.9 FPS to 29.8 FPS, demonstrating its adaptability for low-performance devices. 2025 Elsevier Ltd
Source
Engineering Applications of Artificial Intelligence;Volume;155;Issue;;Article No.;111077;
Date
01-01-2025
Publisher
Elsevier Ltd
Subject
CO2 emissions; Deep-learning; Energy-efficiency; Global warming; Hybrid; Portrait video segmentation
Coverage
Kim Y.-W., Department of Data Science, CHRIST (Deemed to be University) Lavasa, Pune, India; Byun Y.-C., Department of Computer Engineering, Major of Electronic Engineering, Jeju National University, Jeju, South Korea; Krishna A.V.N., Department of Computer Science and Engineering, CHRIST (Deemed to be University), Bangalore, India; Krishnan B., Department of Computer Science and Engineering, CHRIST (Deemed to be University), Bangalore, India
Rights
Restricted Access; Hardcopy may be available in the library
Relation
ISSN: 9521976; CODEN: EAAIE
Format
online
Language
English
Type
Article

Citation

Kim, Yong-Woon; Byun, Yung-Cheol; Krishna, Addapalli V.N.; Krishnan, Balachandran, “High-speed portrait video segmentation using the hybrid combination of deep-learning models and boundary movement adjustment,” CHRIST (Deemed To Be University) Institutional Repository, accessed June 21, 2026, https://archives.christuniversity.in/items/show/22244.

Output Formats