Download color.d and either png.d or bmp.d from my Github repository and put them in your project's directory. In the example, we'll use png.d. Neither module requires any additional D or C libraries.

Let's work with image files by executing the following steps:

The code is as follows:

import arsd.png;
void main() {
        auto image = readPng("test.png").getAsTrueColorImage();
        foreach(ref pixel; image.imageData.colors) {
                int average = (pixel.r + pixel.g + pixel.b) / 3;
                pixel.r = pixel.g = pixel.b = cast(ubyte) average;
        }
        writePng("test-bw.png", image);
}

The following image shows test.png after the conversion:

The modules we dealt with in this recipe implement the file formats independently and provide a simple pixel buffer interface to the image.

The implementation of png.d builds on top of the lazy file input ranges Phobos provides to give a lazy input range of PNG data. This implementation was tricky because the Phobos byChunk range provides a fixed number of bytes, whereas the PNG file format is made out of a stream of variable sized bytes. To bridge the gap, I wrote a BufferedInputRange object with additional consume methods which manage variable sized chunks.

Underneath the readPng and writePng convenience functions, png.d offers both low- and mid-level access to the PNG data structures. The low-level functions provide access to the individual chunks that make up the file, the precise pixel format stored in the file, the palette, and so on. The mid-level functions provide input ranges of PNG file chunks and RgbaScanlines. The range of RgbaScanlines offers a potential hook to use the image data with std.algorithm and other lazy generic transformation functions.