如何获得golang读取jpeg并获得与Python / C相同的unit8值？

Question

I have a test jpeg I'm using for some image processing code & I'm trying to get the same array of uint8 values that opencv in python and C and Pillow (via scipy) in Python, which all match the values I get when I open the jpeg using gimp.

I tried shifting the color values 8 bits, but this does not convert the values accurately - it looks like there's some rounding that I don't quite understand. I know gimp and opencv and Pillow/scipy all use libjpeg, so I tried using https://github.com/pixiv/go-libjpeg and converting the images to 8 bits using

func Convert(img image.Image) *image.RGBA {
   b := img.Bounds()
    rgba := image.NewRGBA(b)
    for y := b.Min.Y; y < b.Max.Y; y++ {
        for x := b.Min.X; x < b.Max.X; x++ {
            r32, g32, b32, _ := img.At(x, y).RGBA()
            c := color.RGBA{uint8(r32>>8), uint8(g32>>8), uint8(b32>>8), 255}
            rgba.SetRGBA(x, y, c)
        }
    }
    return rgba
}

but this still does not match. How can I get the same results?

Answer 1

The JPEG specification doesn't have a pixel exact definition for a decoder, so different implementations will have slightly different output. Even libjpeg on it's own has 4 different Discrete Cosine Transform implementations, using integer or floating point math.

As for your code, if img is already an image.RGBA, the values are all stored as uint8 and you're just shifting the bits back and forth doing nothing. If the image is an image.YCbCr, then you also depend on the color conversion algorithm, and the YCbCr.RGBA() method differs slightly from the color.YCbCrToRGB function. The former tries to maintain more precision, and you need to properly round the values rather than truncating them with a simple right shift. The latter is closer to the spec recommendation, but doesn't use floating point constants which again may differ from other implementations.

Finally, the image itself may have an embedded ICC profile, and depending on the software used to decode the image, the profile may or may not have been applied when viewing the raw pixel data.