CODEWITHSUNDEEP
pythonnumpyimage-processingsvm
SalwoX
SalwoX

Reputation: 1

Centering of numpy image

I want to make an artificial intelligence application that predicts a handwritten number drawn on a canvas based on the MNIST dataset and prints it on the screen. Almost everything is ready for this. My algorithm running successfully in these code but I have a problem. The figure I draw on the canvas is not centered for estimation. I mean, the system cannot detect and predict the number I drew on the corner of the canvas.
These are my codes:

The code I loaded the model

    self.loaded_model = pickle.load(open('svm_model.pkl','rb'))

Used As Canvas Container

    self.label = QtWidgets.QLabel()
    canvas = QtGui.QPixmap(300, 300)
    canvas.fill(QtGui.QColor("black"))
    self.label.setPixmap(canvas)
    self.last_x, self.last_y = None, None

Predict button's function

def predict(self): 
    
    self.label.setAlignment(QtCore.Qt.AlignCenter) 
    s = self.label.pixmap().toImage().bits().asarray(300 * 300 * 4) 
    arr = np.frombuffer(s, dtype=np.uint8).reshape((300, 300, 4))
    arr = np.array(ImageOps.grayscale(Image.fromarray(arr).resize((28,28), Image.ANTIALIAS)))
    arr = (arr/255.0).reshape(1, -1)
    self.prediction.setText('Prediction: '+str(self.loaded_model.predict(arr)))

These are visualizations of my problem:

1

2

I can append all of my codes if you want. I think cause of this problem is not centering the image to be predicted. Because guesses correctly when I draw in the middle of the canvas. I can't find of header in the stackoverflow about my problem. There are similar ones but no matter how hard I tried I couldn't set them to my code.

Upvotes: 0

Views: 440

Answers (1)

norok2
norok2

Reputation: 26886

One way of doing this is by leveraging SciPy's ND-Image function for finding objects in a labelled array scipy.ndimage.find_objects().

Essentially, you would:

  • find the slicing containing the object (assuming the object is whenever the values are different from 0)
  • copy the part of the array within that slicing in the middle of a new array
import numpy as np
import scipy as sp
import scipy.ndimage


def recenter(arr):
    slicing = sp.ndimage.find_objects(arr != 0, max_label=1)[0]
    center_slicing = tuple(
        slice((dim - sl.stop + sl.start) // 2, (dim + sl.stop - sl.start) // 2)
        for sl, dim in zip(slicing, arr.shape))
    result = np.zeros_like(arr)
    result[center_slicing] = arr[slicing]
    return result

To test it actually works:

arr = np.array([
    [0, 0, 1, 0, 0, 0],
    [0, 0, 0, 0, 0, 0],
    [0, 0, 0, 0, 0, 0],
    [0, 0, 0, 0, 0, 2],
    [0, 0, 0, 0, 0, 0],
    [0, 0, 0, 0, 0, 0],
    [0, 0, 0, 0, 0, 0],
    [0, 0, 0, 0, 0, 0],
])

tgt = np.array([
    [0, 0, 0, 0, 0, 0],
    [0, 0, 0, 0, 0, 0],
    [0, 1, 0, 0, 0, 0],
    [0, 0, 0, 0, 0, 0],
    [0, 0, 0, 0, 0, 0],
    [0, 0, 0, 0, 2, 0],
    [0, 0, 0, 0, 0, 0],
    [0, 0, 0, 0, 0, 0],
])

res = recenter(arr)
np.allclose(res, tgt)
# True

Note that in case of mixed parity of the original array shape and the object slicing, the object cannot be placed exactly in the middle and it will be closer to the lower edge along that dimension (when printing these are the top-left edges for 2D arrays).

Upvotes: 1

Related Questions