UID703508
威望3
金钱6074
交易诚信度0
主题23
帖子3940
注册时间2010-6-28
最后登录2024-11-18
中级会员
  
交易诚信度0
注册时间2010-6-28
|
本帖最后由 立華奏 于 2015-11-20 12:39 编辑
各个算法的简单说明:
Nearest Neighbor: A very fast very low quality method, the closest pixel value is used, this looks very bad if not doing integer scaling (1x, 2x, 3x, etc.).
Bilinear: A very fast low quality method. This is as fast as Nearest Neighbor as GPUs can do bilinear filtering in a single operation.
DXVA2: Uses the GPU drivers or fixed function hardware to perform scaling. It is done in 8-bit so it may introduce banding or cause scaled dither noise. It can be very fast and low power. The quality varies between GPU brands and models. Image doubling cannot be used at the same time as DXVA2 scaling. Does chroma upscaling as well so this option overrides 'chroma upscaling'. Requires a restart of madVR to take effect. Does not bypass the graphics card's video (damage) algorithms so make sure they are off.
Mitchell-Netravali: Bicubic with b = 1/3, c = 1/3
Catmull-Rom: Bicubic with b = 0, c = 0.5
Bicubic: Bicubic with b = 0, c = sharpness / 100
SoftCubic: Bicubic with b = softness / 100, c = 0
Lanczos: A sharp sinc type
Spline: A different sinc type
Jinc: A more advanced sinc type (based on an EWA LanczosSharp from ImageMagick)
Bilateral: A chroma scaler that uses the luma channel as reference. This can be very good but it can fail too.
super-xbr: xBR means scale By Rules, it works by detecting edges and interpolating pixels along them. Runs in RGB.
NEDI: New Edge-Directed Interpolation originally proposed by Xin Li and Michael Orchard. Can only perform a 2x upscale. Runs in RGB.
NNEDI3: NNEDI3 (Neural Net Edge Directed Interpolation 3, original by tritical) implemented in OpenCL. Can only perform a 2x upscale but it can double X and Y independently. Below 32 neurons is not recommended for luma because 16 neurons are more likely to generate artifacts by connecting incorrect edges between smaller details. 16 neurons might be fine for chroma as chroma generally lacks small details.
转自其它人:
Nearest Neighbor这是一种最基本、最简单的图像缩放算法,效果也是最不好的,放大后的图像有很严重的马赛克,缩小后的图像有很严重的失真。效果不好的根源在于最临近插值方法引入了严重的图像失真。例如,当反推目标图坐标取得的源图坐标是一个浮点数的时候,采用四舍五入的方法直接获得最接近的象素的值,这种方法是很不科学的。当推得坐标值为0.75时,不应该简单的取为1,目标象素值其实应该根据这个源图中虚拟的点四周的四个真实的点来按照一定的规律计算出来的,这样才能达到更好的缩放效果。双线型内插值算法就是一种比较好的图像缩放算法,它充分的利用了源图中虚拟点四周的四个真实存在的像素值来共同决定目标图中的一个像素值,因此缩放效果比简单的最近邻点插值要好很多。
|
|