Opencv resize changing the Pixel value

Question

While training PixelNet, I have tor resize the annotated image(label) that has specific pixel value,before resizing ;the image pixels has specif values(annotated objects) ,np.unique(image) gives [ 0 7 15] However when i resize the image with openCV to fit it to my network definition,the pixel values range changes,where

image = cv2.resize(image,(cnn_input_size, cnn_input_size),cv2.INTER_NEAREST)

np.unique(bmask) gives [ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17]

This is a disaster for training images with annotated label as these values are specify for other classes,i am wondering if this is the expected behavior of OpenCV while resizing.

Answer 1

Use

image = cv2.resize(image, (cnn_input_size, cnn_input_size), interpolation = cv2.INTER_NEAREST)

or

image = cv2.resize(image, (cnn_input_size, cnn_input_size), 0, 0, cv2.INTER_NEAREST)

Right now you're using the value cv2.INTER_NEAREST for the parameters fx (scale x), and actually using the default interpolation method that is INTER_LINEAR.

Answer 2

The opencv's function of Nearest-Neighbor-Interpolation has bug when the data-type is not UV8. You have to split the data to UV8 types and resize and merge them as follows.

def scale_uint16(img, size):
    img1 = img // 256
    img2 = img % 256
    img1 = cv2.resize(img1.astype('uint8'), size, interpolation=cv2.INTER_NEAREST)
    img2 = cv2.resize(img2.astype('uint8'), size, interpolation=cv2.INTER_NEAREST)
    img3 = img1.astype('uint16') * 256 + img2.astype('uint16')
    return img3