issues with floating point inputs to blob_log, blob_dog, blob_doh

[grlee77]

Description

The blob detection functions all use img_as_float internally to convert integer inputs to normalized floats.

However, if the input is already a float, but has not been normalized, computation can take a very long time.

There is also a separate issue for blob_doh only in that float32 or float16 inputs will result in an error when calling a Cython function that expects double precision.

Way to reproduce

To reproduce, try the docstring example for blob_doh, but casting the input to float. It will get stuck running for a very long time:

from skimage import data, feature
img = data.coins().astype(float)  # non-normalized float
# will get stuck running for a long time
feature.blob_doh(img)

# ValueError: Buffer dtype mismatch, expected 'double_t' but got 'float'
feature.blob_doh(img.astype(np.float32))

Version information

scikit-image version: 0.19.0.dev0

Possible Solution

blob_doh should probably just use img_as_float64 internally so that accuracy is good for the integral_image computation and the input to the Cython function will always be float64.

I assume the slow operation is due to something like a poor default setting for thresholds in the case of non-normalized data. Should we add a check for the input range something like the following, or what is the best way to handle this?

    if image.dtype.kind == 'f':
        if image.max() > 1.0 or image.min() < 0:
            raise ValueError(
                "input image with floating point type must have values normalized to "
                "[0.0, 1.0]")

[odoublewen]

I ran up against this a couple weeks ago. I have unit tests that create "corner case" images, weird artifactural things that should never really happen, but I test them to indentify images that my analysis pipeline can't handle.

Anyway, this simple case will cause blob_log to hang (indefinitely?):

import numpy as np
import skimage
image = np.zeros([100, 100])
image[40:80, 40:80] = 10.0
skimage.feature.blob_log(image, threshold=None, threshold_rel=None)

This is skimage '0.19.0'