HuMomentsDescriptor.py

import numpy as np

from scipy import spatial
from skimage.measure import moments_hu
from image import get_patch


class HuMomentsDescriptor:
    def __init__(self, patch_size=(24, 24)):
        self._patch_size = patch_size

    def extract(self, image, keypoints):
        descriptors = [self.extract_for_keypoint(
            image, kp) for kp in keypoints]
        return np.array(descriptors)

    def extract_for_keypoint(self, image, keypoint):
        patch = get_patch(image, keypoint, size=self._patch_size)
        return self.extract_for_patch(patch)

    def extract_for_patch(self, patch):
        moments = moments_hu(patch)
        # normed_moments = -np.sign(moments)*np.log10(np.abs(moments))
        # - możliwość znormalizowania momentow, ale czysta postac lepiej
        # działa na funkcji distance

        return moments

    def distance(self, desc1, desc2):
        # nie sprawdza sie
        return spatial.distance.cosine(desc1, desc2)

    def distance_correlation(self, desc1, desc2):
        # bardzo slaby
        return spatial.distance.correlation(desc1, desc2)

    def distance_chebyshev(self, desc1, desc2):
        # bardzo duze wartosci, nie sprawdzalem z normalizacja
        return spatial.distance.chebyshev(desc1, desc2)

    def distance_canberra(self, desc1, desc2):
        # troche gorzej jak braycurtis, brak normalizacji
        return spatial.distance.canberra(desc1, desc2)

    def distance_braycurtis(self, desc1, desc2):
        # moj faworyt
        return spatial.distance.braycurtis(desc1, desc2)