img.resize for 4bands image? about remote-sensing-image-semantic-segmentation HOT 4 CLOSED

并且，觉得标签重采样后，边界有一定的问题，如下

from remote-sensing-image-semantic-segmentation.

The method of interpolation sampling generally changes the original data, no matter how many channels the input has. I have already verified this before. The change of the boundary value is in line with the actual situation. Different sampling methods will take values according to the adjacent area when scaling. If the difference between adjacent areas is large, the change of the boundary value will be large. You can think of this as an extra noise.

from remote-sensing-image-semantic-segmentation.

@TachibanaYoshino thansk a lot. I have solved the multiband resizing problem, but I also fould the PlL resize() function only available for uint8 data type. If I want to deal with unint16 data type, the result would be wrong.

from remote-sensing-image-semantic-segmentation.

I have fixed the problem of uint16 type as following:
n_img = np.zeros((dst_h, dst_w, img.shape[-1]), np.uint16)
img = np.asarray(img, np.uint32)
for i in range(img.shape[-1]):
b_img = img[:, :, i]
b_img = Image.fromarray(b_img, mode='I')
b_img = b_img.resize((dst_h, dst_w), mode)
b_img = np.array(b_img, np.uint16)
n_img[:, :, i] = b_img[:, :]

from remote-sensing-image-semantic-segmentation.