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.

what did you expect as a result ? – A. Piro Jun 28 '17 at 07:55 — A. Piro, Jun 28 '17 at 07:55
@A.Piro to keep the same values of pixels [0,7,15] – Eliethesaiyan Jun 28 '17 at 08:08 — Eliethesaiyan, Jun 28 '17 at 08:08

score 12 · Accepted Answer · answered Jun 28 '17 at 08:04

12

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.

answered Jun 28 '17 at 08:04

Miki

40,887
13
123
202

i tried both...give the same result..the i added the zero to see if it doesn't do an scaling – Eliethesaiyan Jun 28 '17 at 08:08
2

the line with `interpolation = cv2.INTER_NEAREST` should work correctly. not sure about the other actually :) – Miki Jun 28 '17 at 08:09
it was due to the fact that i didn't specify the right keyword,this was the problem ecv2.INTER_NEAREST for the parameters fx (scale x), – Eliethesaiyan Jun 28 '17 at 08:15

score 2 · Answer 2 · answered Nov 11 '20 at 08:15

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

Opencv resize changing the Pixel value

2 Answers2