N A T A N . J A C O B S O N
Graduate Student Researcher
Video Processing Lab
ECE @ UCSD

In this paper, we propose an architecture for improving FRUC performance for HD application. The proposed method uses saliency information as well as image segmentation to enhance motion estimation and refinement. Motion vector refinement is applied to salient image regions, while motion field smoothing is applied to non-salient regions. The result is an interpolated sequence of excellent perceptual quality. Objective and subjective results are provided below.

Original Image
3DRS	Full Search
MSEA	Proposed

Subjective Results

Sequence	Method	Std. Dev.	Rej. Region	Average Score
Planes	Prop vs. 3DRS	0.55	0.21	2.24
	Prop vs. FS	1.26	0.49	1.11
	Prop vs. MSEA	0.77	0.30	1.48
Speedway	Prop vs. 3DRS	0.30	0.12	2.81
	Prop vs. FS	0.99	0.38	0.78
	Prop vs. MSEA	1.15	0.44	0.85

Objective Results

Sequence	Metric	3DRS	FS	MSEA	Proposed
Dolphin	PSNR	30.6850	31.8903	31.8539	31.9006
	SSIM	0.9417	0.9504	0.9511	0.9537
Limit	PSNR	37.5492	38.2784	38.5500	38.6604
	SSIM	0.9855	0.9866	0.9871	0.9876
Planes	PSNR	36.6685	37.1436	37.2119	38.2912
	SSIM	0.9940	0.9950	0.9950	0.9952
Speedway	PSNR	25.7847	26.6485	26.6092	26.6846
	SSIM	0.9335	0.9407	0.9404	0.9411

back to main research page