| The widespread use of social networks has generated and accumulated large-scale images.How to store and manage large-scale images efficiently has become an important task of the big data research.In recent years,researchers have proposed to use hashing algorithm to achieve efficient processing of image data.Image hashing algorithm can encode an image of any size into a short sequence of numbers.In practice,the hash sequence can be used to represent the image itself.This can effectively improve computational efficiency and reduce storage space.Currently,image hashing has been successfully applied to retrieval,classification,watermarking,etc.For image data,hashing algorithm usually needs to meet two basic performance metrics,namely robustness and discrimination.Robustness means that,for visually similar input images,the hashing algorithm should encode them into identical or similar hash sequences.In other words,since normal operations such as JPEG compression and brightness adjustment do not change the visual content of an image,the hashing algorithm needs to have the ability to resist these normal operations.Discrimination means that,for those images with different contents,the hashing algorithm should encode them into different hash sequences.Usually,there is a mutual constraint between robustness and discrimination.An increase in one performance often inevitably decreases the other.Currently,most image hashing algorithms have the bottleneck in the classification performance between robustness and discrimination,resulting in limited performance for applications such as copy detection.Therefore,designing efficient algorithms with good balance between robustness and discrimination is an important issue to be addressed in the current image hashing research.Data dimensionality reduction refers to the mapping of data from a high-dimensional space to a low-dimensional space by means of a linear or non-linear mapping function.Generally,data dimensionality reduction techniques have the advantages of reducing redundant information,improving recognition accuracy and maintaining essential structure of the internal data.In this dissertation,four research studies are carried out to investigate efficient image hashing algorithms using data dimensionality reduction techniques such as two-dimensional principal component analysis(2DPCA),two-directional two-dimensional(2D2DPCA)and isometric mapping(Isomap),together with other theories such as visual saliency model,texture feature extraction and polar coordinate transformation(PCT).The first study of this dissertation is an image hashing algorithm based on 2DPCA and feature map(FM),the second study of this dissertation is an image hashing algorithm based on 2DPCA and saliency map(SM),the third study of this dissertation is an image hashing algorithm based on 2D2 DPCA and PCT,and the fourth study of this dissertation is an image hashing algorithm based on Isomap and SM.The main research results of this dissertation are summarized below.1.An image hashing algorithm based on 2DPCA and FM is proposed.2DPCA is a fast and effective technique for data dimensionality reduction,and the use of this technique helps to learn compact image features.Currently,2DPCA has been widely used in image analysis,face recognition,image fusion and object tracking.Since texture features can reflect the spatial and distribution information of an image,the FM construction technique based on texture information can effectively improve discrimination without reducing robustness.Therefore,an image hashing algorithm based on 2DPCA and FM is designed in this dissertation.The specific steps are as follows.Firstly,the local phase quantization features in the frequency domain and the local ternary pattern features in the spatial domain of the pre-processed image are combined to construct an image FM.Secondly,2DPCA is employed to learn low-dimensional features from the image FM.Lastly,the learned features are compressed to generate a hash sequence.Experiments demonstrate that this image hashing algorithm can make a good balance between robustness and discrimination,and its classification performance and copy detection performance are both better than those of many advanced image hashing algorithms.2.An image hashing algorithm based on 2DPCA and SM is proposed.2DPCA can quickly learn low-dimensional and discriminative image features.Since the SM can reflect the saliency region focused by the human visual system,image hash construction with the SM can ensure improvement of robustness.Currently,SM theory has been successfully applied to image quality evaluation,image classification and image retrieval,etc.This dissertation designs an image hashing algorithm based on 2DPCA and SM.The specific steps are as follows.Firstly,SM is calculated by the visual saliency model based on luminance contrast.Secondly,the lowdimensional features are learned from blocks of the SM by 2DPCA.Finally,the distance between the low-dimensional features are computed and quantified to generate a hash sequence.Experiments demonstrate that this algorithm can effectively balance the performances of robustness and discrimination.In addition,this algorithm has a higher computational efficiency than some classical hashing algorithms based on data dimensionality reduction.3.An image hashing algorithm based on 2D2 DPCA and PCT is proposed.Rotation is a common operation of image processing.Identification of similar images attacked by rotation is a challenge problem of the current image hashing research.To address this,a rotation-resistant image hashing algorithm using PCT and 2D2 DPCA is proposed.In this study,the 2D2 DPCA property of translation invariance is firstly theoretically proved.The discovery of this property provides a theoretical basis for the design of a rotation-resistant image hashing algorithm,and is also beneficial to the reseach of other image processing problems.Secondly,a rotation-invariant low-dimensional feature extraction model is designed by combining PCT and2D2 DPCA.Finally,a hash sequence is constructed by using the vector distances of the lowdimensional features output from the 2D2 DPCA.Experiments demonstrate that this algorithm is robust to rotation attack and has good performance in image copy detection under rotation attack.4.An image hashing algorithm based on Isomap and SM is proposed.Since Isomap is a fast and effective non-linear dimensionality reduction method that has been widely used in computer vision,node localization and data visualization.The Isomap can be used to discover geometric features within an image,while the saliency map generated by the frequency-tuned(FT)visual saliency model can reflect the areas focused by the human visual system.And the FT saliency map can be fastly calculated and is robust to noise.In this dissertation,a new image hashing algorithm is designed by combining Isomap and FT saliency map.An important contribution is that the FT visual saliency model is used to generate a saliency map for hash construction,which can ensure the robustness of this algorithm.Another important contribution is that the Isomap is exploited to learn discriminative image features from the saliency map,which can ensure the discrimination of this algorithm.Experiments demonstrate that this algorithm achieves a good classification performance between robustness and discrimination,and outperforms some mainstream image hashing algorithms in image copy detection. |
