Image Hashing Based On Local Invariant Moments And DWT Feature Matrix

Image hashing is a novel topic of multimedia security, and has been widely used in tampering detection, copy detection, digital watermarking, multimedia retrieval, multimedia indexing, and so on. Image hash is visual content based compact representation, which uses a short string of characters/digits to represent an image. If visual content of an image is kept unchanged after digital processing, image hashing should generate the same/similar hashes. This is the first property of image hashing called perceptual robustness, which ensures that hashing can correctly identify those images undergone JPEG compression, watermark embedding, geometric transformation, image enhancement and filtering operations. If two images have different visual contents, image hashing will generate two different hashes. This is the property of discrimination, which can distinguish different images. Moreover, image hashing has additional property in specific applications. For example, image authentication requires security. Therefore, besides the above two properties, image hashing should be key-dependent. In other words, for an input image, different keys will generate different image hashes.This thesis mainly focuses on the techniques of local invariant moments and discrete wavelet transform (DWT), investigates image hashing and its applications in tampering detection and image copy detection, and present two useful hashing algorithms. They are image hashing based on local invariant moments and image hashing based on DWT feature matrix. Detailed research results are as follows.1.1 proposed an image hashing based on local invariant momentsConsidering that invariant moments are invariant to translation, scaling and rotation, I proposed to extract perceptually robust image hashes with local invariant moments from the V component in the HSV color space. Specifically, the input image is firstly pre-processed to obtain the V component of the normalized image. Then, the V component is divided into overlapping blocks. For each block, local invariant moments are extracted to form a feature matrix, which is further conducted data normalization. Next, the normalized matrix is compressed with Euclidean distance, and finally adjacent elements of the matrix are merged and quantified to form image hash. Experimental results illustrate that the proposed algorithm is robust to normal digital operations, such as JPEG compression, watermark embedding, image enhancement, small angle rotation and filtering operations, and has good discriminative capability. To test our capability of tampering detection, an image database with 200 tampered images is constructed. The detection results show that, under a good classification performance, the proposed algorithm can also correctly detect most tampered images, indicating good detection performance.2.1 proposed an image hashing based on DWT feature matrixMost existing hashing algorithms are designed for gray images. For color images, they use luminance components for hash extraction. As color information such as hue and saturation are discarded, their discriminative capabilities are limited. To overcome this problem, I proposed to extract image hash by fully exploiting all components of YCbCr color space. Specifically, input image is firstly preprocessed and converted to Y, Cb and Cr components of YCbCr color space. Then, these components are divided into non-overlapping blocks, and each block is conducted one level DWT. Next, DWT coefficients in four sub-bands are used to construct four feature matrices, which is also performed data normalization. And four normalized matrixes are compressed by L2 norm and two vectors are then obtained. Finally, the two vectors are further compressed by Euclidean distance and quantified to binary string. Receiver operating characteristic (ROC) curve comparisons show that the proposed algorithm is better than some existing algorithms in classification performance between robustness and discrimination. To verify performance in copy detection, a large database with 1210 images is built. The search results show that the proposed hashing has a good detection performance, as well as good classification performance.