Classification of Transformed and Geometrically Distorted Images using Convolutional Neural Network

Abstract

Convolutional Neural Network is a deep learning method that is used in many image-related applications, such as image recognition and classification, it has achieved great performance in these fields, but it still suffers from some shortcomings. One of these shortcomings is not being able to be invariant to the input data due to some image transformations like translation, rotation, scaling, and geometric distortions such as skewness, perspective distortion and pincushion distortion. This study presents an optimized CNN which uses the Geometric Heat Flow (GHF) to improve the performance of the CNN regarding the invariant limitation and classification accuracy. GHF is a partial differential equation that expresses how the heat would diffuse on a surface concerning time in a specific location. GHF is invariant to image transformations and geometric distortions if it was taken concerning the object's arc length which will lead to an invariant CNN. The experiments show that GHF improves the performance of the CNN, and the proposed work achieves an accuracy of 98.09% on the MNIST handwritten dataset, 92.58% on the MNIST-Fashion dataset, and 86.09% on the CIFAR-10 dataset