vincentstimper / mclahe Goto Github PK

Hi, thanks for the code.
I have one dubt. I don't understand the bolded part of the following sentence.

kernel_size: Tuple, list, or numpy array specifying the kernel size along the data dimensions. If kernel_size=None, the kernel size is set to 1/8 of the data size along each dimension. This is a typical choice for photographs. For more complex dataset, the kernel size should be roughly of the size of the features which shall be enhanced by MCLAHE

Hi,

Is there any update on when / whether this will be made compatible with TF2?
I'd be really interested in using your clahe with TF2.

Greetings,
Lydia

Hi Vincent, great library, thank you very much!!! Question, how I can apply mc.mclahe to a color image?

I used the mclahe to enhance my images which size were[128, 128, 80], the super parameter are (n_bins=32, clip_limit=0.01, adaptive_hist_range=Fasle, use_gpu=True). But I found the qualities of images were degraded. Are the super parameters set wrong?

Thanks for the code. Just a few notes in case anyone else runs into the same problems. I needed to add a statement to clear the session when I was using mclahe to loop through thousands of 3D images on a CPU. Otherwise, memory usage expands and code slows to a crawl. Also, the default kernel size is 1/8 of data dimensions. For a 3D image with one dimension less than 8, I got an error. I needed to change kernel size to ceiling of 1/8 image dims. These fixes are shown below.

for i, full_path in enumerate(full_paths):
    os.makedirs(paths_new[i], exist_ok = True)
    print(i)
    img_arr, meta_data = load(full_path)
    dims = np.shape(img_arr)
    img_arr_eq = mc.mclahe(img_arr, kernel_size = [int(np.ceil(dims[0]/8)), int(np.ceil(dims[1]/8)), int(np.ceil(dims[2]/8))])
    save(img_arr_eq, full_paths_new[i], meta_data)
    tf.keras.backend.clear_session()