CODEWITHSUNDEEP
image-processingpython-imaging-library
innuendo
innuendo

Reputation: 373

How to Invert and Normalize a PIL image using the 'point-lambda' function

I was trying to use the point function to invert and normalize a PIL image to 1; but, I am not getting the desired results! What I have tried is this (don't know what is wrong?)

data = data.point(lambda p: 1 if p < 127  else 0 ) # threshold, invert and normalize to 1

For example, when trying 

print(np.array(data).max())

prints True.

However, converting the PIL Image to numpy array and then inverting it worked, as follows: 

data = np.array(data.getdata(),
                np.uint8).reshape(data.size[1], data.size[0], 1)        
        maxG = data.max() # correcting the values of folder e, they do not match the other folders
        data = ( (maxG - data)/maxG ).astype('uint8') 
        tsfm = transforms.ToPILImage() #import torchvision.transforms as transforms
        data = tsfm(data)

I have tried both methods in a word recognition experiment, only the second one worked for me. Sadly, using the point function led to incorrect results. Not sure what is the difference?

NB. The second method is extremely slow, so if this conversion could be done using the point function, that would be a great deal of help.

Upvotes: 1

Views: 660

Answers (1)

Mark Setchell
Mark Setchell

Reputation: 207738

You are confusing "normalisation" and "thresholding".

With "thresholding", you make all values above or equal to a threshold equal to some high number and all values below the threshold equal to some low number. The only possible outcome for each pixel is either the high number or the low number - nothing in between. On a typical 8-bit image, you would threshold at 127, and all pixels would end up as either 0 or 255.

With "normalisation", you constrain all values in the image between some new upper and some new lower limit - or any value between. So the outcome is a bunch of new pixels greater than the low limit and lower than the high limit, but evenly interpolated between the two. On a typical image, you might normalise all values to the range 0-255 and each pixel could end up with any value in that range.

Upvotes: 1

Related Questions