电子学报
電子學報
전자학보
Acta Electronica Sinica
2015年
11期
2218-2224
,共7页
吴少群%袁红星%安鹏%程培红
吳少群%袁紅星%安鵬%程培紅
오소군%원홍성%안붕%정배홍
二维转三维%图割%随机游走%软分割
二維轉三維%圖割%隨機遊走%軟分割
이유전삼유%도할%수궤유주%연분할
2D-to-3D conversion%graph-cuts%random-walks%soft segmentation
半自动2D 转3D 将用户标注的稀疏深度转换成稠密深度,是解决3D 片源不足的主要手段之一.针对现有方法利用硬分割增强深度边缘引入误差的问题,提出像素点与超像素深度一致性约束的边缘保持插值方法.首先,建立像素点深度和超像素深度传播的能量模型,通过像素点与所属超像素间深度差异的约束项将二者关联起来;其次,利用矩阵表示形式将两个能量模型的最优化转换成一个稀疏线性方程组的求解问题.通过超像素提供的约束项,可避免深度传播穿过低对比度边缘区域,从而能保持对象边缘.实验结果表明,本文方法对象边缘处深度恢复的准确性优于融合图割的随机游走方法,PSNR 改善了1.5dB 以上.
半自動2D 轉3D 將用戶標註的稀疏深度轉換成稠密深度,是解決3D 片源不足的主要手段之一.針對現有方法利用硬分割增彊深度邊緣引入誤差的問題,提齣像素點與超像素深度一緻性約束的邊緣保持插值方法.首先,建立像素點深度和超像素深度傳播的能量模型,通過像素點與所屬超像素間深度差異的約束項將二者關聯起來;其次,利用矩陣錶示形式將兩箇能量模型的最優化轉換成一箇稀疏線性方程組的求解問題.通過超像素提供的約束項,可避免深度傳播穿過低對比度邊緣區域,從而能保持對象邊緣.實驗結果錶明,本文方法對象邊緣處深度恢複的準確性優于融閤圖割的隨機遊走方法,PSNR 改善瞭1.5dB 以上.
반자동2D 전3D 장용호표주적희소심도전환성주밀심도,시해결3D 편원불족적주요수단지일.침대현유방법이용경분할증강심도변연인입오차적문제,제출상소점여초상소심도일치성약속적변연보지삽치방법.수선,건립상소점심도화초상소심도전파적능량모형,통과상소점여소속초상소간심도차이적약속항장이자관련기래;기차,이용구진표시형식장량개능량모형적최우화전환성일개희소선성방정조적구해문제.통과초상소제공적약속항,가피면심도전파천과저대비도변연구역,종이능보지대상변연.실험결과표명,본문방법대상변연처심도회복적준학성우우융합도할적수궤유주방법,PSNR 개선료1.5dB 이상.
Semi-automatic 2D-to-3D conversion is a promising solution to 3D stereoscopic content creation.Its main process is to estimate the dense depth map from user-defined strokes on the image.Existing methods preserve depth boundaries by incorpo-rating hard segmentation.However,the inexact segmentation around object boundaries will decrease depth accuracy around these re-gions.To help solve this problem,an edge-aware interpolation method is developed which is constrained by depth consistency be-tween pixels and superpixels.First,we formulate depth propagation in terms of two energy functions of pixels and superpixels,which are influenced by each other through the constraint of soft segmentation.Second,the energy functions are reformulated in matrix forms and they are solved jointly in a sparse linear equation.We recover depth boundaries with help of the superpixels constraint which prevents depth propagation across low contrast edge regions.Experimental comparisons with existing algorithms show that our method demonstrates significant advantages over object boundaries.The PSNR is improved by more than 1.5 dB compared with hy-brid graph-cuts and random-walks approach.