Contour Pattern Recognition with MNIST Dataset

We train a convolutional neural network for handwritten digits from the MNIST dataset. The aim the paper is to test the network for extremal datasets. Actually, we take digits contours as the validating set. The recognition accuracy is expectedly reduced, since the interior of an image carries significant information for the classifier. To obtain the contour of a MNIST -digit, we use the Canny detector and methods of mathematical morphology. The decreasing of the network accuracy may be treated as a part of information which is contained in the interior of handwritten digits.