Convert Numpy Array to Bitmap in C#, Keras.Net

Question

I am working on image semantic segmentation, and trying to figure out how to convert the prediction (after argmax) to bitmap.

If I use ArrayToImg from ImageUtil, it returns a system object. How do I convert it to bitmap from here?

I trained the model using python, and trying to do the prediction in C#.

I tried searching online but so far haven't found any useful information about it.

Answer 1

Updated answer:

Method 1 (NumSharp Method) - Convert Numpy array -> C# array -> NumSharp array -> Bitmap.

//Convert Numpy array to C# array. The Numpy array has shape of (height, width). I use (500,600) for this example:
var array_csharp = array_numpy.GetData<long>();

/Convert C# array to NumSharp, and reshape it to back to image size
NumSharp.NDArray gray_numsharp = array_csharp;
gray_numsharp = gray_numsharp.reshape(500, 600, 1);

//create an empty RGB array with white color background
NumSharp.NDArray rgb_numsharp = NumSharp.np.ones((500, 600, 3)) * 255;

//Convert class index to color. Note the color value is in BGR order instead of RGB
for (int row = 0; row < 500; row++)
    for (int col = 0; col < 600; col++)
        if (gray_numsharp[row, col, 0] == 1)
        {
            rgb_numsharp[row, col] = new int[] { 0, 0, 255 };
        }
        else if (gray_numsharp[row, col, 0] == 2)
        {
            rgb_numsharp[row, col] = new int[] { 255, 0, 0 };
        }

//Reshape into 4 dimensions, required for bitmap conversion
rgb_numsharp = rgb_numsharp.reshape(1, 500, 600, 3);

//Change datatype to byte
NumSharp.NDArray numsharp_byte_array = array_rgb_numsharp.astype(NumSharp.np.@byte);

//Convert to bitmap and save it
Bitmap mask = numsharp_byte_array.ToBitmap();
mask.Save(@"C:\prediction.png", ImageFormat.Png);

Method 2 (without using NumSharp)

• Convert Numpy array -> C# array -> Bitmap. I took reference from here. I am also using the function MySaveBMP from the link.

//Convert Numpy array to C# array. The Numpy array has shape of (height, width). I use (500,600) for this example:
var array_csharp = array_numpy.GetData<long>();

int width = 600;
int height = 500;
int bytesPerPixel = 3;
int pixel_length = width*height;
byte[] bgr_array = new byte[pixel_length*3];

for (int i=0; i < pixel_length; i++)
{
    if (array_csharp[i] == 1)
    {
        bgr_array[3 * i] = 0;
        bgr_array[3 * i + 1] = 0;
        bgr_array[3 * i + 2] = 255;
    }
    else if (array_csharp[i] == 2)
    {
        bgr_array[3 * i] = 255;
        bgr_array[3 * i + 1] = 0;
        bgr_array[3 * i + 2] = 0;
    }
    else
    {
        bgr_array[3 * i] = 255;
        bgr_array[3 * i + 1] = 255;
        bgr_array[3 * i + 2] = 255;
    }
}

MySaveBMP(bgr_array, width, height);

Both the methods were tested and work as expected. I would choose to use method 2 due to the speed. Method 1 takes like 2 sec (from Numpy array to bitmap conversion), while method 2 is almost instantly I can get the result.

---

Below is earlier answer and can be ignored

After researching here and there, I think this might work, though I have not done a complete test yet. Put it might help someones who have the same question:

1. First convert the numpy array to C# array using GetData:

var c_sharp_array = numpy_array.GetData<int>();

2. Then convert it to NumSharp and reshape it to desired shape:

NumSharp.NDArray numsharp_array = c_sharp_array;
numsharp_array = numsharp_array.reshape(x,y,z);

3. Convert to Bitmap using NumSharp library:

Bitmap bmp = numsharp_array.ToBitmap();

Will mark this as accepted answer once it is fully tested.