Hybrid RGB-D Segmentation

Hak Cipta

Hybrid RGB-D Segmentation

Hybrid RGB-D segmentation is an innovation in image segmentation that integrates three key approaches - mathematical optimization: statistical modeling, and deep learning - into a single, integrated pipeline. The key innovation of this method is the direct synergy between the Active Contour (Snake) algorithm, Poisson Matting, Closed-Form Matting, and Convolutional Neural Network (U-Net), which work together to produce object segmentation with high precision and extremely smooth boundaries. This system automates trimap formation using a combination of dynamic contours and depth information, thereby eliminating the need for manual annotation, a common bottleneck in the matting process. The innovative use of depth data via K-Means clustering facilitates object isolation based on spatial differences, resulting in high robustness against complex backgrounds with similar colors. The implementation of Just-In-Time (JIT) acceleration in Poisson Matting is another technical innovation that improves computational efficiency without compromising accuracy. The multi-method integration yields a stable, adaptive, and realistic alpha mask. This program makes a significant contribution to the development of precise, automated RGB-D segmentation with a hybrid approach that is not yet available in similar systems, making it ready for application in various advanced computer vision applications.

2026-1776401153-qakn

B-8444/III.6.4/TK.11.01/4/2026


( Lihat )

Pusat Riset Kecerdasan Artifisial dan Keamanan Siber

puji.lestari@gmail.com

Badan Riset dan Inovasi Nasional

SF - Perangkat Lunak

Bandung

14 April 2026

EC002026070023

22 Mei 2026

22 Mei 2026

001243147

  • Puji Lestari
    ( Pusat Riset Kecerdasan Artifisial dan Keamanan Siber )
  • Elli A. Gojali
    ( Pusat Riset Kecerdasan Artifisial dan Keamanan Siber )
  • Mukti Wibowo
    ( Pusat Riset Kecerdasan Artifisial dan Keamanan Siber )
  • Lidya Ananda Talalu
    ( Universitas Pertahanan Republik Indonesia )
  • Muhammad Haikal Ziaulhaq
    ( Universitas Pertahanan Republik Indonesia )
Kembali