Image Hashing Algorithms Based On Low-Rank Models

Wide applications of the Internet provide massive images.Consequently,storage and management of massive images have become important issues to be addressed.Image hashing technology generates a short number sequence based on image content for representing image.The digital sequence is called image hash,which greatly improves the efficiency of image storage and transmission and provides an efficient technique for handle massive images.Image hashing can not only be applied to image copy detection and image authentication,but also to many other application scenarios,such as social media event detection,image retrieval,and image quality assessment,etc.Robustness and discrimination are the most basic properties that image hashing algorithms should satisfy.Robustness means that the image hashing algorithm must map visually similar images to the same or similar hash sequences no matter whether the file data representations of the images are same or not.In other words,the image hashing algorithm should be robust to normal digital operations(e.g.,compression,filtering,and enhancement).This is because digital operations change the file data of an image but do not change its visual content.Discrimination means that the image hashing algorithm must map different images to completely different hash sequences.Since the number of different images is much larger than the number of similar images in practice,good discrimination means that the error rate of judging different images as similar ones is low.Obviously,these two properties are mutually constrained,so designing high-performance algorithms that take both into account is an important task in current image hashing research.In this paper,we use low-rank models such as low-rank representation and low-rank sparse matrix decomposition,and also combine with theories and techniques such as visually salient models,ring partition and invariant vector distance to study image hashing algorithms.The first research work is to design the image hashing algorithm jointly using low-rank representation and ring partition,and the second research work is to design the image hashing algorithm jointly using low-rank sparse matrix decomposition and invariant vector distance.The main research results are as follows.1.Image hashing algorithm based on low-rank representation and ring partition is proposed.Low-rank representation is a useful technique to obtain the global structure of data,which is robust to noise and can extract the lowest-rank representation of all data.The theory of low-rank representation has been widely used in many fields,such as subspace segmentation,image segmentation and image classification.In this paper,a novel image hashing algorithm is proposed jointly with low-rank representation and ring partition.The algorithm obtains the visual saliency map through the spectral residual saliency model and uses it to construct a weighted image representation of the preprocessed image.Then,the low-rank representation technique is applied to the weighted image representation,and the rotation-invariant hash is extracted from the lowrank representation by ring partition.The hash algorithm is verified with extensive experiments,and the experimental results show that the hash algorithm can achieve a good balance between robustness and discrimination,and its area under the receiver operating characteristic curve is larger than that of some state-of-the-art hashing algorithms.2.Image hashing algorithm based on low-rank sparse matrix decomposition and invariant vector distance is proposed.The low-rank sparse matrix decomposition can decompose the image into a low-rank matrix and a sparse matrix,which can effectively obtain the image features.The low-rank matrix contains the basic structural features of the image and can describe the image content well.Therefore,this paper designs a novel image hashing algorithm using low-rank sparse matrix decomposition and invariant vector distance.The algorithm uses low-rank sparse matrix decomposition to obtain the low-rank matrix,and the low-rank matrix is partitioned into non-overlapping blocks,and extracts the statistical features of each blocks to form a feature vector.Finally,the invariant vectors distances are calculated using Euclidean distance to form a compact image hash.The robustness and discrimination of the algorithm are verified on two open datasets,and the experimental results show that the algorithm achieves a good balance between robustness and discrimination,and outperforms several comparative literature algorithms.