Bilinear Resize Implementation (Python)

bilinear_resize.py

def sample(img, x: float, y: float):
    h, w = img.shape

    left = math.floor(x)
    top = math.floor(y)
    right = left + 1
    bottom = top + 1

    # top left corner
    if (0 <= left < w) and (0 <= top < h):
        a = img[top, left]
    else:
        a = 0

    # top right corner
    if (0 <= right < w) and (0 <= top < h):
        b = img[top, right]
    else:
        b = 0

    # bottom left corner
    if (0 <= left < w) and (0 <= bottom < h):
        c = img[bottom, left]
    else:
        c = 0

    # bottom right corner
    if (0 <= right < w) and (0 <= bottom < h):
        d = img[bottom, right]
    else:
        d = 0

    # linear interpolation of top two points
    top_interleaved = (right - x) * a + (x - left) * b

    # linear interpolation of bottom two points
    bottom_interleaved = (right - x) * c + (x - left) * d

    # linear interpolation of top and bottom points
    return (bottom - y) * top_interleaved + (y - top) * bottom_interleaved


 def saturate(value, low, high):
    if value < low:
        return low
    if value > high:
        return high
    return value


def get_source_index(
    i: int, scale: float, low: int, high: int, align_corners: bool = False
):
    if align_corners:
        return scale * i
    else:
        return saturate((i + 0.5) * scale - 0.5, low, high)


def bilinear_resize(x: np.ndarray, size: Tuple[int, int], align_corners: bool = False):
    y = np.empty(size, dtype=x.dtype)

    if align_corners:
        scale_h = (x.shape[0] - 1) / (y.shape[0] - 1)
        scale_w = (x.shape[1] - 1) / (y.shape[1] - 1)
    else:
        scale_h = x.shape[0] / y.shape[0]
        scale_w = x.shape[1] / y.shape[1]

    for i in range(y.shape[0]):
        for j in range(y.shape[1]):
            i_src = get_source_index(i, scale_h, 0, x.shape[0] - 1, align_corners)
            j_src = get_source_index(j, scale_w, 0, x.shape[1] - 1, align_corners)
            y[i, j] = sample(x, j_src, i_src)

    return y

TODO

It is still not exactly aligned with OpenCV when data type is UINT8.

Reference

• PyTorch's Upsampling Source Code