Classification of images by using deep learning methods based on perceptual hash functions

Machine learning, image classification, video analysis, speech recognition and natural language processing are widely used by data scientisits in both industrial and academic world for various purposes. The manuel extraction of the features requires a certain expertise during the image classification stage. Thus, the feature extraction phase is considered to be the most important step in pattern recognition for image classification. Deep learning algorithms have become an active research area to remove the manuel feature extraction step of images. However, image processing take considerable long execution times as it operates at the pixel level. The salient characteristics of images are determined by perceptual hash functions which give a certain hash value. Therefore, the feature data obtained by the perceptual hash function can also be called the fingerprint of the image. In this thesis, salient features are obtained directly from raw images by using perceptual hash function which does not distort image structure. Comparisons have been made by considering some parameters as BER, MSE, PSNR to evaluate the performance of the proposed perceptual hash function. Liver disease images obtained from Elazig University Hospital Radiology Laboratory, and Caltech-101 (publicly available database) images were used as a database. The Convolutional Neural Network (CNN) algorithm, one of the deep learning architectures, was used with the Discrete Wavelet Transform (DWT) - Singular Value Decomposition (SVD) based perceptual hash function. The aim of this doctorate thesis is to reduce the classification time of the high-dimensional images and their sizes on hard disk while maintaining the classification performance above an acceptable threshold by using perceptual hash based CNN. It is intended to extract the features by using both DWT - SVD based perceptual hash function and the CNN for the first time together. The results obtained by applying the our proposed method to the liver and the Caltech-101 image database were evaluated by considering some known classifiers such as Artificial Neural Networks (ANN), Support Vector Machines (SVM), K-nearest Neighbor (KNN). The proposed method not only improves the classification performance but also provides a positive impact on the excessive runtime problem.