Research On Image Set Compression With Sparse Dictionary

Image is the main source of information.However,with the development of image acquisition devices,the amount of image data acquired by people is increasing rapidly,which brings great difficulties to image storage and transmission.Therefore,it is of great significance to seek efficient image/video compression technology.In daily life,a large number of images are obtained from the same or similar scene,and they are usually organized and stored as an image set.There is some information redundancy between the images in the set,and the efficient use of these redundant information will improve the efficiency of the image set compression and save the storage space.Therefore,how to make full use of the image set redundancy information based on the traditional single image compression technology is one of the main research directions of image set compression.Traditional image compression usually obtains a compact representation of the signal by transform,and then completes the image coding process through quantization and entropy coding.For example,the most widely used JPEG standard uses a DCT substrate to transform the signal.Fixed substrate can not adapt the signal content characteristics.In addition,some complex signals can not be represented efficiently and we can not get a more compact representation.In this paper,we introduce the sparse representation theory and the traditional image compression technology and propose a method of image set compression based on sparse dictionary.The main research contents include the following two aspects:Firstly,an exponential distribution sparse dictionary model is proposed,and an image set compression scheme based on this dictionary is established under the traditional image coding framework.Sparse representation can be more flexible and efficient to represent complex signals compared with DCT and other fixed substrates,and it can adapt to the content characteristics of the image set.In this paper,a new sparse representation model based on exponential distribution is proposed.When the signal is approximated on the dictionary,the coefficient not only has sparse characteristics,but also can show a certain exponential distribution decline law,and thus has certain compressible characteristics.The experimental results show that,at low bit rate the proposed method is better than the traditional JPEG method in both subjective quality and objective quality.Secondly,a content adaptive image set compression scheme based on the regression least squares(RLS)sparse dictionary is proposed.Compared with the traditional sparse representation Dictionary(such as the K-SVD dictionary),the RLS dictionary reduces the dictionary’s dependence on the training data by constantly updating the training samples and introducing the forgotten factor in the iterative process,making the dictionary training results independent of the initialization dictionary.The RLS dictionary has the stability characteristics as an offline dictionary,thus reducing the cost of dictionary transmission.In this paper,we use the RLS dictionary algorithm to realize the purpose of adaptive image set content by using a fixed external image database and classifying training data.Dictionary classification makes the non-zero sparse coefficient distribution more compact,as well as the energy is more concentrated,which makes the compression efficiency further improved.In addition,the decoder uses the image non-local similarity feature and the external image database similarity feature to enhance the reconstruction quality.The experimental results show that the proposed method can effectively improve the image set compression performance.