VipsOperation: class "new_from_memory" not found

[stuta]

I'm using twain to read raw data from scanner but vips.Image.new_from_memory() prints:
error: 'no such operation
VipsOperation: class "new_from_memory" not found
'

local depth = twain.TWAIN_DibDepth(hdib) --  Depth of DIB, in bits i.e. bits per pixel.
local width = twain.TWAIN_DibWidth(hdib) -- Width of DIB, in pixels (columns)
local height = twain.TWAIN_DibHeight(hdib) -- Height of DIB, in lines (rows)
local rowBytes = twain.DIB_RowBytes(hdib) -- Number of bytes needed to store one row of the DIB.
local size = height * rowBytes
local data = ffi.new("unsigned char[?]", size) -- unsigned char[?]== BYTE[?]
local dst
for nRow = 0, height - 1 do
 dst = data + (nRow * rowBytes)
 twain.DIB_ReadRow(hdib, nRow, dst) -- Read row n of the given DIB into buffer at dst
end
local image = vips.Image.new_from_memory(data, size, width, height, depth, C.VIPS_FORMAT_UCHAR) -- C.VIPS_FORMAT_UCHAR == 0

[jcupitt]

Hello @stuta, lua-vips is missing new_from_memory -- I just never needed it.

It should be simple to add -- here's the equivalent in node-vips, it's just 8 lines:

https://github.com/jcupitt/node-vips/blob/master/lib/image.js#L253

I'll tag this as an enhancement.

[jcupitt]

I had a quick go in a branch. It seems to work for me -- would you be able to look it over and test it? Example:

            local width = 64
            local height = 32
            local size = width * height
            local data = ffi.new("unsigned char[?]", size)
        
            for y = 0, height - 1 do
                for x = 0, width - 1 do
                    data[x + y * width] = x + y
                end
            end

            local im = vips.Image.new_from_memory(data, width, height, 1, "uchar")

            assert.are.equal(im:width(), width)
            assert.are.equal(im:height(), height)
            assert.are.equal(im:bands(), 1)
            assert.are.equal(im:format(), "uchar")
            assert.are.equal(im:avg(), 47)

We should probably add write_to_memory as well.

[jcupitt]

git master has picked up quite a few improvements now. There should probably be a new release.

[stuta]

Thanks, vips.Image.new_from_memory() works in osx and win7/x86.

I'm having trouble to understand how do I create picture data from uchar array. For example if picture depth is 1 and picture width is 34 pixels then uchar array size is 5 bytes (34 / 8 bytes of uchar = 4.25 = 5 bytes).

In my scanner test I get picture: width = 1275, depth = 24 and uchar array size = 3828 (that can hold up to 1276 pixels width). Picture height = 1753 does not matter here because I get 1 row uchar array at a time.

[jcupitt]

Great! I'll have a go at adding write_to_memory() as well.

RGB images are three bytes per pixel (or 24 bits), so you need something like:

            local width = 64
            local height = 32
            local sizeof_line = 3 * width
            local data = ffi.new("unsigned char[?]", height * sizeof_line)
        
            for y = 0, height - 1 do
                for x = 0, width - 1 do
                    data[y * sizeof_line + x + 0] = x + y
                    data[y * sizeof_line + x + 1] = x + 2 * y
                    data[y * sizeof_line + x + 2] = 2 * x + y
                end
            end

            local im = vips.Image.new_from_memory(data, width, height, 3, "uchar")

[stuta]

Thanks, code below works when bit depth is >= 8 (greyscale = 8 bit). How about b&w 1 bit images? Depth 1/8 did not work.

local depth = twain.TWAIN_DibDepth(hdib) --  Depth of DIB, in bits i.e. bits per pixel.
local width = twain.TWAIN_DibWidth(hdib) -- Width of DIB, in pixels (columns)
local height = twain.TWAIN_DibHeight(hdib) -- Height of DIB, in lines (rows)
local rowBytes = twain.DIB_RowBytes(hdib) -- Number of bytes needed to store one row of the DIB.
local dst = ffi.newNoAnchor("unsigned char[?]", rowBytes)
local byteDepth =  math.floor(depth / 8)
local byteCount = width * byteDepth
local data = ffi.newNoAnchor("unsigned char[?]", height * byteCount) -- unsigned char[?] == BYTE[?]
for y = 0, height - 1 do
  twain.DIB_ReadRow(hdib, y, dst) -- Read row n of the given DIB into buffer at dst
  for x = 0, width - 1 do
    for z = 0, byteDepth - 1 do
      data[(y * byteCount) + (x * byteDepth) + z] = dst[(x * byteDepth) + z]
    end
  end
end
local image = vips.Image.new_from_memory(data, width, height, byteDepth, "uchar")

[jcupitt]

libvips does not support one-bit images in memory. You need to unpack to one pixel per byte, with 0/255 for black/white. It can pack down to one bit for (eg.) tiff save.

[stuta]

I tried this kind of packing, but it produces garbage:

byteDepth = 1
data = ffi.newNoAnchor("unsigned char[?]", height * 8 * rowBytes) -- unsigned char[?] == BYTE[?]
for y = 0, height - 1 do
  twain.DIB_ReadRow(hdib, y, dst) -- Read row n of the given DIB into buffer at dst
  for x = 0, rowBytes do
    for z = 0, 7 do
      data[(y * rowBytes) + (x * 8) + z] = dst[x]
    end
  end
end

Is there a way to set dpi info to picture when saving it? Now 400dpi scanned documets open as huge pictures.

[stuta]

Finally I have working code.

local depth = twain.TWAIN_DibDepth(hdib) --  Depth of DIB, in bits i.e. bits per pixel.
local width = twain.TWAIN_DibWidth(hdib) -- Width of DIB, in pixels (columns)
local height = twain.TWAIN_DibHeight(hdib) -- Height of DIB, in lines (rows)
local rowBytes = twain.DIB_RowBytes(hdib) -- Number of bytes needed to store one row of the DIB.
local byteDepth =  depth / (ffi.sizeof("unsigned char") * 8)
local byteCount = width * byteDepth
util.printInfo("used scanner width: %d, height: %d, bit depth: %d, rowBytes: %d, twain state: %d", width, height, depth, rowBytes, state)
local dst = ffi.newNoAnchor("unsigned char[?]", rowBytes)
local data
if byteDepth >= 1 then -- grayscale or color 
  data = ffi.newNoAnchor("unsigned char[?]", height * byteCount) -- unsigned char[?] == BYTE[?]
  for y = 0, height - 1 do
    twain.DIB_ReadRow(hdib, y, dst) -- Read row n of the given DIB into buffer at dst
    for x = 0, width - 1 do
      for z = 0, byteDepth - 1 do
        data[(y * byteCount) + (x * byteDepth) + z] = dst[(x * byteDepth) + z]
      end
    end
  end
else -- less than 8 bytes per pixel
  local pixelsPerByte =  math.floor(8 / depth)
  byteDepth = 1 -- vips image can't have less than 1 pixel per byte
  if math.floor(byteCount) < byteCount then
    byteCount = math.floor(byteCount) + 1
  end
  data = ffi.newNoAnchor("unsigned char[?]", height * width)
  for y = 0, height - 1 do
    twain.DIB_ReadRow(hdib, y, dst) -- Read row n of the given DIB into buffer at dst
    for x = 0, byteCount - 1 do
      for z = 0, pixelsPerByte - 1 do
        --[[ -- long version
        if dst[x] > 0 and dst[x] < 255 then
          opt = bit.lshift(1, pixelsPerByte - z - 1)
          val = bit.band(dst[x], opt)
          val = val > 0 and 255 or 0
        else
          val = dst[x]
        end --]]
        if (x * pixelsPerByte) + z < width then
          data[(y * width) + (x * pixelsPerByte) + z] = bit.band(dst[x], bit.lshift(1, pixelsPerByte - z - 1)) > 0 and 255 or 0
        end
      end
    end
  end
end

[jcupitt]

Nice!

I added :write_to_memory() as well, though there's an extra copy in there I couldn't get rid of, unfortunately.

        local psize = ffi.new(gvalue.psize_typeof, 1)
        local vips_memory = vips.vips_image_write_to_memory(self.vimage, psize)
        local size = psize[0]
        -- can we avoid the copy somehow?
        local lua_memory = ffi.new(gvalue.mem_typeof, size)
        ffi.copy(lua_memory, vips_memory, size)
        glib.g_free(vips_memory)

Do you know of a way to avoid that? You'd think you could make a cdata object from a pointer and length, then attach the g_free() to it with ffi.gc().

new_from_memory() should probably allow a string as well as a ffi memory area.

[stuta]

This is how I just created picture read and conversion for zbar barcode recognition. It works perfectly in osx x64 and win7 x86.

You should try that ffi.gc(), I have not used it much.

Is there a 64-bit version windows libvips?

local image = vips.Image.new_from_file(path)
local interpretation = image:interpretation()
if interpretation ~= "b-w" then
  image = image:colourspace("b-w") -- srgb -> b-w
end
image = image:flatten()
local raw = image:write_to_memory()
return image:width(), image:height(), raw