Performance test of Sws_scale algorithm in FFmpeg

The Sws_scale in FFmpeg is often used for image scaling and format conversion, and the function can use a variety of different algorithms to process the image. Used to be lazy, lazy to test and identify which algorithm should be used, the most recent work time, many times need to wait for others. Sneak in, the ffmpeg of this set of functions were encapsulated, by the way, testing a variety of algorithms.

Briefly speaking about the test environment, I am using the Dell brand machine, i5 CPU. FFmpeg was compiled from the latest version around August 2010, and I used its static library version.

The Sws_scale algorithm has the following options.

#define SWS_FAST_BILINEAR     1
#define SWS_BILINEAR          2
#define SWS_BICUBIC           4
#define Sws_x                 8
#define SWS_POINT          0x10
#define SWS_AREA           0x20
#define Sws_bicublin       0x40
# Define Sws_gauss          0x80
#define SWS_SINC          0x100
#define Sws_lanczos       0x200
#define Sws_ SPLINE        0x400

First, a 1920*1080 landscape image is scaled to the 400*300 24-bit RGB, and the following frame rate refers to the number of times per second scaled and rendered. (After my tests, the rendering time is negligible, and the main time is spent on the scaling algorithm.) ）

Algorithm	Frame rate	Subjective perception of images
Sws_fast_bilinear	228	The image has no obvious distortion and feels very good.
Sws_bilinear	95	It feels good, smoother than the edge of the previous algorithm.
Sws_bicubic	80	Feel the same, more than the algorithm edge to smooth, than the previous algorithm to be sharp.
Sws_x	91	With the previous image, I don't see the difference.
Sws_point	427	The details are sharper and the image effect is slightly less than the above figure.
Sws_area	116	I don't see the difference with the algorithm.
Sws_bicublin	87	Ditto.
Sws_gauss	80	In relation to the previous algorithm, some of the smoothing (or blurring) is possible.
Sws_sinc	30	The details should be clearer relative to the previous algorithm.
Sws_lanczos	70	For the previous algorithm, to be smooth (or fuzzy) a little bit, almost no difference.
Sws_spline	47	And the previous algorithm, I do not see the difference.

Overall, the above various algorithms, the image after the effect seems to be good. If you are not looking at the contrast, you can hardly see the effect of scaling. The above-mentioned clarity (sharpness) and smoothing (blurring), is an objective feeling, not clear is better than smooth, and not smooth than clear good. The point algorithm, the efficiency of high, let me shock, but the effect is not bad. In addition, I compared the use of CImage drawing when the zoom, its frame rate can be 190, but the effect is terrible, color serious distortion.

The second experiment, the image of a 1024x768 landscape, zoom to 1920*1080, and rendering (at this time, although not negligible, but not more than 5ms rendering time, does not affect the relative accuracy of the following conclusions).

Algorithm	Frame rate	Subjective perception of images
Sws_fast_bilinear	103	The image has no obvious distortion and feels very good.
Sws_bilinear	100	There is no difference from the above figure.
Sws_bicubic	78	Relative to the picture, feel the details a little bit clearer.
Sws_x	106	No difference from the picture above.
Sws_point	112	The edges are visibly jagged.
Sws_area	114	The edges have no obvious sawtooth.
Sws_bicublin	95	There is almost no difference from the above picture.
Sws_gauss	86	A little clearer than the edge of the above picture.
Sws_sinc	20	No difference from the picture above.
Sws_lanczos	64	No difference from the above figure.
Sws_spline	40	No difference from the above figure.

Overall, the point algorithm has obvious sawtooth, area algorithm sawtooth is not obvious, the rest of the algorithm, the naked eye does not appear to be significantly different. In addition, when rendering with CImage, the frame rate can be scaled up to 105, and the effect is similar to point.

Personal advice, if the image of the scaling, to pursue efficient, for example, video image processing, in the unclear is to zoom in or zoom out, directly using the sws_fast_bilinear algorithm can be. If you explicitly want to zoom out and display, it is recommended to use the point algorithm, if it is clear to zoom in and display, in fact, the use of CImage strech more efficient.

Of course, if the speed of the pursuit of picture quality. In the above algorithm, the selection of the lowest frame rate, the screen effect is generally the best.

But overall, the ffmpeg scale algorithm, the speed is very fast, after all, I choose the material is high-definition pictures.

(I would like to upload a picture, but the different groups of pictures are actually very small, I am afraid that when the upload format conversion caused by the loss of image details, has exceeded the details of each picture itself, so here do not upload pictures.) )

Note: Test the resize efficiency of the OPENCV, and the same situation above, OPENCV in the above amplification test, can be carried out 52 times per second, reduce the test, can be carried out 458 times per second.

Original address: http://www.cnblogs.com/acloud/archive/2011/10/29/sws_scale.html

Updated (2014.8.5) ================

FFmpeg Command Example (Bilinear,bicubic,neighbor) using a different Sws_scale () scaling algorithm:

Ffmpeg-s 480x272-pix_fmt yuv420p-i src01_480x272.yuv-s 1280x720-sws_flags bilinear-pix_fmt yuv420p Src01_bilinear_12 80X720.YUV
ffmpeg-s 480x272-pix_fmt yuv420p-i src01_480x272.yuv-s 1280x720-sws_flags bicubic-pix_fmt yuv420p src 01_BICUBIC_1280X720.YUV
ffmpeg-s 480x272-pix_fmt yuv420p-i src01_480x272.yuv-s 1280x720-sws_flags neighbor-pix_ FMT yuv420p SRC01_NEIGHBOR_1280X720.YUV