The Study Of Image/Video Signal Processing Methods In The Compressed Domain Based On Compressed Sensing

The sampling mechnisms for images/videos based on compressed sensing(CS)can take advantages of the inherent characteristics of compressed sampling.By using CS,there is no need to adopt the frequency that is twice the maximum frequency of the original signal and then throw away plenty of useless coefficients during the compression process as the traditional sample methods,which causes the waste of resources.Therefore,compared with the traditional sampling,compression and transmission algorithms for images/videos,CS has some overwhelming advantages,such as the reduction in sampling data,the low sampling complexity and so on.Based on this,CS has wide applications in image aquisition,medical images,channel estimation,wireless sensor networs,etc.Recently,the research on the recovery algorithms for CS is almost sufficient.However,applictions like signal recoginition,image classification do not involve the reconstruction process.They can accomplish the expected goals only by using measurements.This kind of methods targeting this issue is called compressive signal processing(CSP).The advantages of CSP methods lie in that we can obtain enough characteristics of the original signal directly from the measurements which have less data amount.Note that the measurements are not equal to the original signal,they can not be used directly as some signal processing methods in the pixel domain.As a result,there is a necessity to explore the relationship between the compressed domain and the pixel domain in order to reveal the information of the original signal contained in the measurements.In this way,we can perform image/video processing directely in the compressed domain.Based on the CS theory,this dissertation tries to develop the image/video processing methods in the compressed domain.By analyzing the relationship between the measurements and the original signal,we directly perform image/video processing with measurements,such as texture images classification,motion estimation and saliency analysis.In specific,the main four contributions are as follows.1)In the first part,we study the relationship between the measurements in compressed domain and coefficients in frequency domain based on CS.In specific,we take the measurements and frequency coefficients of image blocks as samples to construct thecross-covariance matrix in the compressed domain(CCCD)and cross-covariance matrix in the transform domain(CCTD).By theoritical analyses,we find that they have a linear relationship and this kind of relationship improves with the increase of the number of samples and the measurement rate.Then,by using the texton dictionary which consists of the covariance vectors,we can directly perform texture classification algorithms with favourable accuracy and speed.2)In the second part,we study the motion estimation methods in the compressed domain.Since we only have the measurements of non-overlapping blocks whose positions are relatively fixed after the one-time compressive sampling,it is quite difficult to obtain the measurements of block at any position in one frame.To solve this issue,we first establish the spatial relationship between the measurements of the non-overlapping blocks and the block to be estimated.We decompose the motion vector into the horizontal direction and the vertical direction,then we use the measurements of the non-overlapping blocks in the reference frame to estimate the measurements of the current block.In this way,compared to the method which directly takes the colocated block in the reference block as the matched block,more accurated one can be found to improve the reconstruction performance of images and videos.3)In the third part,we study the saliency of images and videos based on CS.Since saliency has a close relationship with the engergy distribution,we expore the saliency of images and videos in terms of the DCT transform and K-SVD in the compressed domain.Firstly,combining with the Parseval theorem,we find that the 2-norms of the DCT coefficients in the pixel domain approximate that of coefficients in the compressed domain.Furthermore,after analyzing the DCT equation,we conclude that each frequency component of the DCT coefficients in the compressed domain is the weighted sum of the components of the original signals.Since the energy of the signal can be more concentrated by using the dictionary trained by K-SVD method,we propose the compressive K-SVD dictionary training method to remove the redundancy parts while preserve the sparse ones in order to generate the saliency map.The proposed method can be used to extract salient features in the media cloud.This feature extraction method has advantages over CNN in both speed and storage.4)In the forth part,the distributed compressed video sensing architecture based on regionsaliency and the applications of DCVS in the vehicular infotainment systems are studied.The proposed distributed compressed video sensing architecture has the codec moduals for the key frames and non-key frames,saliency judgement,twice sampling,quantization and dequantization moduals.Then the applications in the vehicular infotainment,such as safe driving,privacy preserving,vehicular crowdsensing,vehicular communications,traffic estimation are discussed.Lastly,we demonstrate the applications of DCVS in the vehicular infotainment.By taking advantage of the relationship between the measurements from the sampling devices in the vehicular network,we can use the decoded frames to help the reconstruction of the subsequent frames.Simulation results show that the proposed DCVS system has a superior video reconstrucion quality compared with the conventional DCVS system.