Reduce number of tiles of an image-CodePudding

I have a list of images that I want to simplify by generating a tile list and indexing the image.

Right now I'm testing with a single image and trying to find out a good alogorithm to compare the tiles. I might have to throw everything out and see if I can do something better. Right now I have this code:

def tileize(img):

 tile_list = []

 new_img = np.zeros((int(IMGX/TILESIZE),int(IMGY/TILESIZE),0))

 i = 0
 ri = 0
 for r in range(0,img.shape[0],TILESIZE):
    cj = 0
    for c in range(0,img.shape[1],TILESIZE):
        t1 = img[r:r TILESIZE, c:c TILESIZE,:]
        new_img[ri,cj] = i
        tile = Tile(t1)
        tile_list.append(tile)
        i =1
        cj  =1
    ri  =1

return tile_list,new_img

I then would like to use the tile_list and "shrink" it. I thought to create Tile objects so that I can create a set()

class Tile(object): 
 def __init__(self, mat):
    self.mat = mat
    self.hist = cv2.calcHist([mat],[0],None,[256],[0,256])
 def __eq__(self, other):
    res = cv2.compareHist(self.hist,other.hist,cv2.HISTCMP_CORREL)
    if res > 0.95:
        return True
    return False
 def __hash__(self):
    return hash(self.hist.flatten())

Which doesn't work because hist is an array.If I just compare the Tile object together (combine) I won't get a shorter list.

What is the best method to reduce the tile_list by comparing its items?

CodePudding user response：

You should not use __eq__, __hash__ and set like this since they are not designed for what you are looking for, i.e. approximate equality. One fondamental property of __eq__ and __hash__ is that if two objects are equal then their hash should be equal too. This invariant is not respected in your implementation.

If you have a small number of patches, you could just compare each patch with every other patches and then group them by similarity. Take a look at clustering algorithms for that purpose.

Another option closer to what you already tried is to use Perceptual Hashing algorithms. Take a look at the ImageHash Python library and the examples. The example from find_similar_images.py seems close to your application.