Histogram equalization in fastai

Part 1 (2020)
1

Does fastai2 has integrated transformation for histogram equalization? If no, what would be the easiest way to implement it as custom transformation and pass it to batch_tfms in ImageLoader?

2

Since nobody offered any solution here is mine. It is slow, but works:

# Parts taken from: https://medium.com/hackernoon/histogram-equalization-in-python-from-scratch-ebb9c8aa3f23

# create our own histogram function
def get_histogram(image, bins):
    # array with size of bins, set to zeros
    histogram = np.zeros(bins)
    
    # loop through pixels and sum up counts of pixels
    for pixel in image:
        histogram[pixel] += 1
    
    # return our final result
    return histogram

# create our cumulative sum function
def cumsum(a):
    a = iter(a)
    b = [next(a)]
    for i in a:
        b.append(b[-1] + i)
    return np.array(b)

class HistogramEqualization(Transform):
    def __init__(self):
        return
    def encodes(self, o):
        # convert our image into a numpy array
        imgnp = o.cpu().numpy()  # np.asarray(o[0]) # .permute(1, 2, 0))

        # put pixels in a 1D array by flattening out img array
        flat = imgnp.flatten()
        # execute our histogram function
        hist = get_histogram(flat, 256)

        # execute the fn
        cs = cumsum(hist)
        # numerator & denomenator
        nj = (cs - cs.min()) * 255
        N = cs.max() - cs.min()

        # re-normalize the cumsum
        cs = nj / N

        # cast it back to uint8 since we can't use floating point values in images
        cs = cs.astype('uint8')

        # get the value from cumulative sum for every index in flat, and set that as img_new
        img_new = cs[flat]

        # put array back into original shape since we flattened it
        img_new = np.reshape(img_new, o.shape)

        ret = TensorImage(img_new) if (type(o) == TensorImage) else o
        return ret
        
    def decodes(self, o):
        return o
1 Like
3

Faster solution is with item_tfms instead od batch_tfms:
from fastai.vision.core import PILImage
class HistogramEqualization_item(Transform):
def init(self, prefix=None):
self.prefix = prefix or “”
def encodes(self, o):
if type(o) == PILImage:
ret = PIL.ImageOps.equalize(o)
else:
ret = o
return ret
def decodes(self, o):
return o
data = ImageDataLoaders.from_folder(path, train=‘train’, valid=‘test’,
item_tfms=[Resize(size=384)], HistogramEqualization_item()],
batch_tfms=[*aug_transforms(min_scale=0.98, do_flip=False)],
max_zoom=1.1, max_lighting=0.2, bs=16, num_workers=8)

2 Likes