Image Hash Searching

Abstract

In the big data generation, the quantity of images has increased sharply in the computer world. Duplicate and similar images are becoming a problem. To deal with the problem, image hashing provides a fast way to compare images. In this project, two different hashing skills are used. Firstly, Perceptual Hash Algorithm is a common skill to hash photos, which includes average, different and perceptual hashing. Secondly, due to the development of deep learning, the Convolution Neural Network (CNN) is a good way to deal with image classification by extracting features from a multiple convolution layer. By adding a hidden layer between the output features and the classification layer, the hash code can be extracted from the hidden layer. Indeed, the hash code includes the image features and part of the semantic meaning. By combining the two hashing methods, an image search image engine and a duplicate photo detector can be built, which can retrieve images rapidly compared to computing the Euclidean distance between image features. Two different tests have been conducted. Their results show that the bit length affects the computation time and the image retrieval rate depends on the accuracy of the CNN model. Furthermore, the system is rotation invariance.