ImageDataGenerator 中的图像数组与 CV2 读取的图像不匹配

Question

I use the Image Data Generator with flow from directory to read in a single 224 X 224 X 3 image titled 1.jpg. The result should be and identical to the initial image since image size is set to 224 X 244 in flow from directory. I then read in the same image using cv2. I then compare the array provided by the image data generator to the array produced by the cv2 read image. I expected them to be identical but there are NOT. The saved images from the generator and that saved by cv2 look identical when viewed. To check I printed the first 10 values in each array and they differ. Why? Code is shown below.

dir=r'c:\Temp'# path to  directory containing the sub directory imgtest. 
# sub directory imgtest contains one directory called class1, it contains the image file 1.jpg which is already shape 224 X 224 X 3
test_dir=r'c:\Temp\imgtest' # path to test images directory
save_dir=dir # where the generator will storge the image on disk so it can be viewed later
test_gen=ImageDataGenerator().flow_from_directory(test_dir,
                target_size=(224, 224), batch_size=1, class_mode='categorical',color_mode='rgb',save_to_dir=save_dir,save_format='jpeg' ,shuffle=False )
data=test_gen.next() # get the next batch from the generator -will be only 1 file which is 1.jpg
image1=data[0][0] # this is the single image 1.jpg provided by the test_gen
print ('for image provided by the generator image shape is ', image1.shape,  '  data type is ',image1.dtype )
img_dir=r'c:\Temp\imgtest\class1'
img_path=os.path.join(img_dir, '1.jpg')
img=cv2.imread(img_path,cv2.IMREAD_UNCHANGED) # read in the original 1.jpg image from c:\Temp\imgtest\class1
img = img.astype('float32') #cv2 reads in data as int8 so convert to float32 to match generator data type
write_loc=os.path.join(dir, 'cvimage.jpg') 
cv2.imwrite(write_loc, img) # save the image to disk so it can be viewed later
print ('for image read in by cv2 image shape is ', img.shape, ' data type is ', img.dtype)
compare_arrays = (image1 == img).all()
if compare_arrays:
    print('arrays are the same')
else:
    print('arrays do not match')
print ('Generator Data,     CV2 Data       Delta')
for i in range (0,10):
    delta=image1[i][0][0] - img [i][0][0]
    print( '  {0}              {1}         {2}'.format(image1[i][0][0],img [i][0][0], delta ))
# the results if running the code are shown below
Found 1 images belonging to 1 classes.
for image provided by the generator image shape is  (224, 224, 3)   data type is  float32
for image read in by cv2 image shape is  (224, 224, 3)  data type is  float32
arrays do not match
Generator Data,     CV2 Data       Delta
  129.0              155.0         -26.0
  141.0              164.0         -23.0
  156.0              174.0         -18.0
  165.0              180.0         -15.0
  173.0              182.0         -9.0
  179.0              183.0         -4.0
  180.0              181.0         -1.0
  181.0              178.0         3.0
  185.0              176.0         9.0
  180.0              170.0         10.0
The original image, generator image and cv2 image are shown in the composite image
[![composite of original, generator and cv2 saved images][1]][1]





  [1]: https://i.stack.imgur.com/TCDDp.jpg

Answer 1

The issue here was that openCV's standard method for reading images is to read them as BGR, so that blue is the first channel and red is the last. However, most other standard libraries will use RGB (keras's ImageGenerator and PIL.Image, as examples), and so you cannot directly compare cv2 objects with these objects, as they will be different mathematically.

However, after saving to a picture file, this issue should not be an issue, as formatting like png is standard.