performance: vision.AlphaBlender is slower when image is big.

vision.AlphaBlender is a very good function that can do a lot of "image matting", but when I encounter a large image size, the speed is very slow, look forward to the official enhancement of the efficiency of this function.

blender = vision.AlphaBlender('Operation','Binary Mask',...
    'MaskSource','Input port');

% big image test performance
HH = [8000,500];
WW = [8000,500];
for i = 1:length(HH)
    H = HH(i);
    W = WW(i);
    
    bottomImg = zeros(H,W,3,'single');

    topImg = rand(H,W,3,'single');
    maskImg = zeros(H,W,'logical');
    maskImg(1:100,1:100)=1;

    t1 = tic;
    outImg = blender(bottomImg,topImg,maskImg);
    t2 = toc(t1);
    fprintf('image size:(%d*%d) take time: %.3f seconds\n',H,W,t2)
end

image size:(8000*8000) take time: 2.445 seconds
image size:(500*500) take time: 0.148 seconds

Answer 1

[xingxingcui]

use imblend build-in function instead of vision.AlphaBlender since R2024b.

------------------original answer-----------------

There are many ways to solve this, such as rewriting the implementation of this function yourself in C/C++ and wrapping it as a mex. I now think it's not just about improving the algorithm, it's about external factors adjusting to my own needs, even if I rewrite it in an optimised C/C++ (not taking into account margin effects), such as cv::copyTo, they don't really differ much in performance, see the performance curve below.

note: uint8 type, two dims image array test