Improving the CNN algorithm using a novel hybrid method

There are many different types of neural networks today, but the convolution neural network is one of the most popular one. This network is very popular due to feature extraction from images, videos, etc. In this paper, we first apply three fundamental changes to the convolution neural network architecture and thus introduce a new convolution neural network that is very resistant to noise. Then we compare the newly introduced algorithm. We do this for the MNIST dataset in noisy and non-noisy mode. The results show that even if we add ۴۰% noise to the original data, the output of the proposed method is the same as the none-noise mode.We then suggest using the IMCNN + KNN hybrid algorithm to increase the classification accuracy. For this purpose, we use the ABIDE۱ database related to Magnetic Resonance Imaging of Autism Spectrum Disorder (ASD).The accuracy of classifying Normal Control with autism in the proposed method, even in the presence of noise, is ۹۹.۴%, which is a significant improvement over the CNN algorithm.

Keywords:

Autism Spectrum Disorder (ASD) , improved convolutional neural network (IMCNN) , k-nearest neighbors algorithm (KNN) , Noise reduction