Image loses quality with cv2.warpPerspective

Question

I am working with OpenCV 3.1 and with Python.

My problem comes when I try to deskew (fix the tilt of) an image with text. I am using cv2.warpPerspective to make it possible, but the image loses a lot of quality. You can see here the original part of the image:

and then, here, the "rotated" image:

it is like smoothed.

I was using morphological transformation like:

kernel = np.ones((2, 2), np.uint8)
blur_image = cv2.erode(tresh, kernel, iterations=1)

and

white_mask2 = cv2.morphologyEx(white_mask2, cv2.MORPH_OPEN, kernel)

to see if it improves something, but nothing.

I saw this example here in SO, but those guys had the same problem: and

So, I don't have idea what can I do. Maybe there is a way to not losing the quality of the image, or, there's another method to rotate the image without quality lost. I know this method:

root_mat = cv2.getRotationMatrix2D(to_rotate_center, angle, 1.0)
result = cv2.warpAffine(to_rotate, root_mat, to_rotate.shape, flags=cv2.INTER_LINEAR)

But it doesn't work for me, as I have to rotate every rectangle here:

and not the whole image. It means, the best way I found to do it, was warpPerspective, and it works fine, but with quality loss. I would appreciate an advice to avoid the quality lost.

Answer 1

The problem is related to the interpolation required by the warping. If you don't want things to appear smoother, you should switch from default interpolation method, which is INTER_LINEAR to another one, such as INTER_NEAREST. It would help you with the sharpness of the edges.

Try flags=cv2.INTER_NEAREST on your warp call, be it warpAffine() or warpPerspective().

Interpolation flags are listed here.

enum InterpolationFlags { 
    INTER_NEAREST = 0, 
    INTER_LINEAR = 1, 
    INTER_CUBIC = 2, 
    INTER_AREA = 3, 
    INTER_LANCZOS4 = 4, 
    INTER_MAX = 7, 
    WARP_FILL_OUTLIERS = 8, 
    WARP_INVERSE_MAP = 16 
}