A New Method for Classification of Images Using Convolutional Neural Network Based on Dwt-Svd Perceptual Hash Function

Abstract

This paper proposes a method by using Convolutional Neural Network (CNN), which reduces the image classification time and maintains the classification performance above an acceptable threshold. A hybrid model called Discrete Wavelet Transform Singular Value Decomposition based Perceptual Hash Convolutional Neural Network (DWT-SVD-PH-CNN) is proposed by using a perceptual hash function together with CNN to reduce the classification time. In the proposed method, the DWT-SVD- based perceptual hash function is used. The most important feature of perceptual hash functions is to obtain the salient features of images. First, DWT-SVD based perceptual hash function is applied to images for obtaining salient features. Then, images making up of salient features, are produced in 32 x 32 format and given as inputs to CNN, where Support Vector Machine (SVM) is used to classify the images. In this paper, the DWT-SVD-PH-CNN method is applied to Caltech 101 image database. Experimental results show that the proposed DWT-SVD-PH-CNN method has a high accuracy, about 95.8 %. Moreover, this method reduces the execution time from 241.21 seconds to 83.08 seconds compared to the classical method. Thus, the experimental results show that the proposed DWT-SVD-PH-CNN method performs much faster than classical CNN by maintaining the image classification accuracy high.