juglab / n2v

This is the implementation of Noise2Void training.
Other
392 stars 107 forks source link

Input images have to be 32-bit #43

Closed faymanns closed 4 years ago

faymanns commented 4 years ago

The input image to model.predict(img, axes='YX', n_tiles=(2,1)) has to be a 32-bit image. Using 16-bit images results in hot pixels in the background with values close to the saturation value of 16-bit.

tibuch commented 4 years ago

Hi @faymanns

This is true, the networks expect 32-bit float images. We should probably add a test and convert-step in our predict method.

Thanks for reporting!