faster lean bilinear imresize / improved gpuArray/imresize

Hi,

I'm currently processing lots of images in a convolutional neural network, and for that the resize.m function is currently the major bottleneck (Googling this resulted in a few other people complaining about imresize as well). However, digging a bit into the code there is about 60% overhead for this function by checking arguments and performing extra function calls. So I made a leaner version, but this requires access to the private imresizemex function. Is it possible in a future release to, for example:

• create a lean imresize_bilinear function (as attached here)?
• move imresizemex outside of the private directory so it can be accessed? (I work on several different servers, often with a different matlab version, so copying imresizemex does not work too well for me).

Related:

• Are you working on making gpuArray/imresize work in format: out = imresize(im, [numRows numCols])?

Example code is attached. Output:

   >> testImresize
   original bilinear resize: 2.364755
   lean bilinear resize: 0.922745

Thanks, Jasper