How to Flatten and stack Efficiently 50000 Images

Question

I have 50000 grayscale images with 56*56 pixels each one. I need to flatten images and stakck it in the same array with dimensions 50000*3136 after that export the file as .CSV. I am doing this code which load all images in the memory which can cause a problem if we have not enough memory space.

imagesToArray[pathsrc_, pathdes_] := Module[{filesList, data}, (
   (* pathsrc_ : list of images path *)
   (* pathdes_ : CSV file path*)

   filesList = FileNames["*.png", pathsrc];
   Print["Number of images :", Length@filesList];
   data = 
    Table[Flatten[ImageData[Import[filesList[[i]]]]], {i, 1, 
      Length@filesList}];
   Export[pathdes, data];
   data
   )]

Is there any alternative to do it faster without memory problem?

Answer 1

You can combine @nikie's approach above with the use of a faster built-in import function.

If you can forgo all the behind-the-scene checking that Import does, you can try using the built-in function that Import ultimately calls to read in a PNG file, which is Image`ImportExportDump`ImageReadPNG. I found that out by using Trace@Import[somePNGfile] and wading through quite a lot of output. Incidentally, the counterpart to this function also exists, i.e. ImageWritePNG, but it requires multiple arguments and I haven't figured that one out yet; however, Export is already quite a bit zippier than Import at least for PNG files, so the need is less pronounced there.

It returns a list containing the image imported. To get each image, you can use e.g.

First@Image`ImportExportDump`ImageReadPNG[pathToImage]

It is at least 10x faster than using Import, and quite possibly more.

For instance, let's generate 200 small PNGs of the kind you are working with:

MapIndexed[
  Export["images\\" <> ToString[First@#2] <> ".png", #1] &,
  Image /@ RandomReal[{0, 1}, {200, 56, 56}]
];

Now let's compare timings:

First@Image`ImportExportDump`ImageReadPNG["images\\" <> ToString[#] <> ".png"] & /@ 
   Range[1, 200]; // AbsoluteTiming

Import["images\\" <> ToString[#] <> ".png"] & /@ Range[1, 200]; // AbsoluteTiming

(* Out:
{0.221601, Null}
{15.5932, Null}
*)

That's a 70x speedup!

Answer 2

You can use stream functions, like this (untested):

f = OpenWrite[pathdes];
chunkSize = 100;
Monitor[Do[
  (
     lines=ParallelTable[Flatten[ImageData[Import[filesList[[j]]]]],
            {j,i,Min[i+chunkSize-1,Length@filesList]}];
     WriteString[f, ExportString[lines, "CSV"]];
  ), 
  {i, 1, Length@filesList, chunkSize}], i];
Close[f];

to avoid the memory cost. I'd guess Import would be the bottleneck in that code, and there's not much you can do to make it faster.