Idea
Rencently CNNs reache the state-of-the-art in the image classification or recognition tasks. We use convolutional kernels to extract image features for classification.
Hough Transformation (HT for short) performs well in parameteric shapes such as lines, circles or eclipses. Therefore the author wants to utilize the property of Hough Transformation (HT for short) in CNN to improve the classification or recognition accuracy.
Network Structure
This paper proposes a CNN with two batches which both are a convolution layer with a fully-connected layer flowed. One gets original images as input. The other gets images through HT progress as input.
My Opinion
It's not HT neural net because it just ues HT as a preprocess method. This combination is actually mechanical and not used in the first time.
Shall we find a way to change the HT filter to a convolutional kernel? And is it helpful to improve the CNN?
