OpenCV Image Depth is Corrupted by JSON

Question

I am working on a project which requires the transfer of OpenCV frames using JSON, where the frames are used in a CascadeClassifier on the receiving end. I have encountered the error in the CascadeClassifier:

Kyles-MBP:facial_detection kyle$ python3 badcv.py
Traceback (most recent call last):
  File "badcv.py", line 30, in <module>
    doesnt()
  File "badcv.py", line 26, in doesnt
    faces = cascade.detectMultiScale(imgnew)
cv2.error: OpenCV(3.4.2) /Users/travis/build/skvark/opencv-python/opencv/modules/objdetect/src/cascadedetect.cpp:1376: error: (-215:Assertion failed) scaleFactor > 1 && _image.depth() == 0 in function 'detectMultiScale'

Kyles-MBP:facial_detection kyle$

I have distilled my code to isolate the error below. Clearly I am doing something wrong, but I do not have enough experience with OpenCV to know what to do here. I did some searching, and an image depth 0 corresponds to CV_8U, however I have no idea how to set such a depth (I did some searching and concluded that it shouldn't matter, since cv2 natively represents images as ndarray, but this could be a false assumption). Furthermore, I cannot identify any difference between the pre and post jsonified ndarray; by my estimate, on all measures save for physical location in memory, the pre and post arrays are identical. I have included interpreter outputs below from investigating the data structures.

What am I doing wrong, and how do I avoid encountering this particular error? Thanks!

Code:

# badcv.py
import cv2
import json
import numpy as np
import os
import sys

cascade_file_path = os.path.dirname(
    os.path.realpath(__file__)) + '/default.xml'

def works():
    img = cv2.imread(sys.argv[1])
    imgnew = img

    rows, cols = imgnew.shape[:2]
    cascade = cv2.CascadeClassifier(cascade_file_path)

    faces = cascade.detectMultiScale(imgnew)

def doesnt():
    img = cv2.imread(sys.argv[1])
    data = { 'file': json.dumps(img.tolist()) }
    imgnew = np.array(json.loads(data['file']))

    if not (img is imgnew):
        print("Not the same object")

    rows, cols = imgnew.shape[:2]
    cascade = cv2.CascadeClassifier(cascade_file_path)

    faces = cascade.detectMultiScale(imgnew)

if __name__ == "__main__":
    works()
    doesnt()

As an addendum: The default.xml file is the Haar xml classifier that ships with OpenCV and I am using a simple test file that is 10px X 20px, however this script still fails on all sizes of images, across jpg and png.

Equality:

I also checked for object equality, and the following assertion is valid:

>>> if (img == imgnew).all(): print("element-wise equality)
'element-wise equality'

However object level equality is not valid (which makes sense because json.loads will return a new dictionary, not a cached one in memory):

>>> if not (img is imgnew): print("not the same object")
'not the same object'

The types of both img and imgnew are ndarray with the same shape:

>>> if type(img) is type(imgnew): print("same type")
'same type'
>>> type(img)
<class 'numpy.ndarray'>
>>> if img.shape == imgnew.shape: print("same shape")
'same shape'

Answer 1

I did some searching, and an image depth 0 corresponds to CV_8U, however I have no idea how to set such a depth

You were on the right track there. This is the bit depth of the image, the data type of each pixel. img will be loaded with a dtype of np.uint8, an unsigned 8-Bit integer, which is the same as CV_8U.

When you pass through json, the pixel values become Python integers, and the created numpy array will have a np.int64 dtype.

Thus the issue:

>>> img.dtype == imgnew.dtype
False

Can be rectified with:

# Create an array with 8-Bit unsigned integers
imgnew_u8 = imgnew.astype(np.uint8)