Image Hashing Algorithms Based On Locally Linear Embedding And Locality Preserving Projection

Image hashing is an important research problem in the cross field between information security and image processing, which has been successfully used in many applications, such as content authentication, image forensics, watermark embedding and tamper detection. Actually, it is a new technology of image representation, which can map any size image into a short sequence of characters/bits. In general, digital image is often undergone some digital manipulations in the practical use. After these digital manipulations, image data is changed, but visual content of image is kept unchanged. Therefore, image hashing should map images with similar visual contents to the same or similar hashes. This property is called robustness of hashing algorithm, which can ensure that image hashing can correctly identify those images generated by digital operations, such as brightness adjustments, contrast adjustments and JPEG compression. The second property of image hashing is uniqueness. It means that images of different visual contents should be converted to significantly different hashes. This is helpful to distinguish different images accurately. Besides the two properties mentioned above, image hashing has other property in specific applications. For example, in application to image forensics, it should be enough secure to avoid malicious tamper and forgery. Generally, image hashing algorithm can be divided into two steps, where the first step is image feature extraction and the second step is feature compression and coding. In fact, data reduction is a useful technology for feature compression and coding. It can map high-dimentional data to low-dimentional space by linear or non-linear mapping, so as to discover the low dimensional structures hidden in the high-dimensional observations. And the results obtained can be used in feature compression and coding.This thesis mainly focuses on two classic dimensionality reduction methods (i.e., locally linear embedding (LLE) and locality preserving projection (LPP)), investigates data reduction based image hashing algorithms, and obtains two interesting results. They are LLE-based image hashing, and LPP and Gabor filter based image hashing. Detailed research results are illustrated as follows.1.I proposed an image hashing algorithm based on statistical features of LLE.LLE can reflect local geometry structure of non-linear manifold and compactly describe data relation. Here,I investigate the use of LLE in image hashing, and find the property that embedding vector variances of LLE are approximately linearly changed by normal digital operations. Based on this observation, I propose a novel LLE-based image hashing. Specifically, an input image is firstly mapped to a standard size and then converted to CIE L*a*b* color space, where the L* component is used to construct a secondary image. Finally, LLE is applied to the secondary image and the embedding vector variances of LLE are used to form image hash. Hash similarity is determined by correlation coefficient.2.I proposed an image hashing algorithm based on LPP and Gabor filter.LPP, Gabor filter and chaotic map are exploited to construct image hashing algorithm. Firstly, input image is converted to a fixed size by bilinear interpolation. If the input image is a color image, it is converted to YCbCr color space and the luminance component Y is used for representation. Then, image blocks are used to construct a sencondary image, which is filtered by a Gabor filter. Next, LPP is applied to the secondary image for data reduction, and variances of low dimensional vectors are extracted to form intermediate hash. Finally, chaotic map is exploited to conduct data encryption and the final image hash is obtained. Hash similarity is evaluated by Hamming distance.Many experiments are conducted to validate robustness and discrimination of the above proposed algorithms. The experimental results show that the proposed algorithms are both robust against normal digital processing, such as brightness adjustment, gamma correction, image scaling, and salt & pepper noise contamination, and have good discrimination for distinguishing images with different contents. Receiver operating characteristics (ROC) curve comparisons with some hashing algorithms are also done, and the results illustrate that the proposed algorithms are better than the compared algorithms in classification performance.