基於邊緣散焦程度的單相機影像快速深度估測方法

Abstract

本論文提出了一個針對單視域影像的深度估測方法，為了從單張影像獲得深度資訊，我們使用模糊線索作為估測依據。模糊線索估測的方式都是根據影像由光學系統成像時，物體離焦平面越遠，因散焦造成的模糊程度就會越嚴重，換個角度來說就是該物體的成像會越平滑。在影像中，越平滑代表的是訊號強度變化率越低，也就是高頻成分越少，大部分的模糊線索估測方法也都是依據該區域高頻成分多寡進行深度估測。 但是直接使用高頻能量大小當作估測依據將會使得估測結果容易受到該區總能量的大小影響，也就是說環境亮度以及物體顏色將會影響估測的結果。另外，我們在研究的過程中也發現其實模糊後的高頻能量不是消失，而是由高頻轉移到低頻。所以我們提出了一個基於高低頻能量比例的估測方法來解決因為亮度及顏色不同而造成的估測誤差，並對運算複雜度進行優化。 實驗中，我們使用由視差線索估測出的深度圖當作真實深度(Groud Truth Depth Map)，將模糊線索估測出的深度圖與真實深度圖計算出PSNR與SSIM進行比較。結果證實，本論文提出的方法在高亮度情況下可以在像素對像素的相關性指標(SSIM-Structure)中獲得30.97%的提升。在中低亮度時，其他方法為負相關，本論文提出的方法依然保持正相關。在運算速度上則減少了65%的運算時間。

This paper presents a method for depth map estimation from monoscopic video. To estimate depth from single image, we use defocus cue in our method. Defocus depth map estimation method uses lens blur to estimate depth map. If an object is not on the focal plane, it will be blurred. A blurred image is smoother then original. It means we can calculate blurred level by spatial variety or energy in high frequency. Most of the defocus depth map estimation method use the total energy in high frequency to estimate the depth map. But the problem is that total energy in high frequency will effect by luminance and color in the image. It means that the estimation result will be effected by different luminance and color. Then we find out that high frequency energies will not disappear but move into low frequency part in a blurred image. Based on this theory, we proposed a method that estimating depth map from ratio of high and low frequency, and we optimize the computational complexity of this method. We use disparity depth map as ground truth depth map, and compare it to defocus depth map to calculate PSNR and SSIM. Experiment results show proposed method is more stable in different color and luminance. In high luminance case, proposed method can improve 30.97% at SSIM-Structure comparison. In low and medium luminance case, other method get negative SSIM-Structure value, and our method is still positive. Proposed method only need 65% less computational time in speed comparison.