Study On Compressed Sensing Based Image Processing And Applications

With the rapid development of multimedia technology and the increasing popularity of various intelligent electronic devices,multimedia communication,especially,images and video data have gradually replaced text as the mainstream form of the modern communication,which is accompanied with higher requirements for the quality and security of images.However,the image signal is different from ordinary text or audio data.It has a large amount of data,high redundancy,and strong pixel correlation.How to accurately and securely transit and reconstruct images in the field of multimedia communication becomes a hotspot.Compressed sensing is able to compress and sample signal simultaneously.The amount of transmitted data is much smaller than that sampled by the traditional theory.In the reconstruction process,the high-dimensional image can be recovered from a small number of low-dimensional projection measurements accurately with high probability by solving the optimization problem.The application of this theory significantly reduces the cost of data acquisition,storage and transmission,and saves signal processing time and equipment cost.CS provides a theoretical basis for image processing in multimedia communication applications.Based on CS theory,this thesis studies the classified dictionary construction method with the aim of realizing the image transmission and ensuring the low-complexity image acquisition.Based on this dictionary construction,the super-resolution method is designed.Reconstruction algorithm and image processing method for secure transmission,and image classification based on measurement domain features are studied.The specific work of this thesis is summarized as follows:Firstly,based on CS,the correlation between the measurement domain and the frequency domain is analyzed.The relationship model between these two domains is established.The characteristics of the measurement domain can reflect the structural features of the image pixel domain.Then we classify the image blocks based on the correlation characteristics of the measurement domain,and divide them into smooth,texture and edge blocks according to their structural characteristics.This classification algorithm based on the measurement domain has been proved to be able to classify image blocks accurately and efficiently.Finally,based on the classification of image blocks,we update the dictionary atoms by solving the optimization problem on each type of image block sample set and train a redundant dictionary that can sparsely represent the image blocks.At the same time,we use other methods to train the dictionary and compare them with our proposal.It is concluded that the measurement domain classification dictionary can make full use of the structural information of the image blocks,and retain the detailed information of some images so that the image signal can be more sparsely represented.Furthermore,we also reconstruct the test image by using the constructed classification dictionary,and compare it with other algorithms.The experiment shows that the quality of the reconstructed image based on the classification dictionary is higher than other algorithms and with lower algorithm complexity.The PSNR values in our proposal are increased by about 1.64-6.7d B.The reconstruction time is shortened as one-quarter of traditional DCT method and decreased by 1.8-3s compared with the GMSDL method.Secondly,we propose an image super-resolution(SR)method based on CS and the inherent similarity of an image to recover the high-resolution(HR)image from a single lowresolution(LR)image.Because the difference of image blocks is not considered when training dictionaries,a dictionary classification method based on measurement domain is proposed.Specifically,we use the linear relationship among images in the measurement domain and frequency domain to classify image blocks into smooth,texture and edge feature blocks in the measurement domain.The dictionaries for different blocks are trained using different categories.Consequently,an LR image block of interest may be reconstructed using the most appropriate dictionary.If one merely learns the prior knowledge from the external image database,it tends to generate untrue details of the reconstructed HR image.In our proposed method,we use the nonlocal similarity of the image to tentatively search for similar blocks in the whole image and present a joint reconstruction method based on CS and similarity constraints.The sparsity and self-similarity of the image blocks are taken as the constraints.This method performs better both visually and quantitatively than some existing methods.The PSNR value of the recovered image is increased by about 0.56-3d B compared with the existing algorithm,and the SSIM is increased by 0.01-0.11.Thirdly,an image reconstruction method for secure image transmission is proposed.The measurement process of CS has quite strong secrecy performance.The encryption method based on CS theory decreases resources demand during signal acquisition and protect the image data to avoid theft of confidential information.Moreover,the CS-based encryption method can achieve the encryption and compression of an image simultaneously.This thesis proposes an encryption scheme based on a single-round dictionary and chaotic sequences.The measurement matrix and the random shuffle of the locations of the measurements depend on two different Logistic sequences.In addition,the starting values are treated as elements of the key.Equally significant,our proposal quantifies the measurements such that they are easily transmitted and stored.The secret key is composed of three parameters derived from the logistic sequences and scrambling process.Therefore,the scheme noticeably reduces the resources required to transmit the secret key.Moreover,by using a single-round dictionary as the substitute for the DCT basis,each image has its own unique dictionary,allowing the image to be adaptively encrypted and reconstructed.This feature leads to a significant enhancement of the security level and drastic improvements in the quality of the decrypted image.Both the experimental and analytical results verify the proposal’s effectiveness,high level of security and substantial image quality improvement after decryption.The key space is approximately 1040,and the correlation coefficients of the adjacent pixels in the ciphertext image can be as low as 0.0108.The quality of the reconstructed image is also improved,and the PSNR value is increased by 0.5-3.8d B over the existing algorithms.Lastly,CS theory is applied into intelligent traffic system for vehicle detection and classification.It mainly consists of two parts.First,CS is used to generate a saliency map in the measurement domain for window calibration.Then,the CS measurement process is embedded into the traditional Convolutional Neural Network(CNN)as the compression layer,and the CS-CNN network is obtained.Then we use the CS-CNN to classify the target vehicles in the traffic scene and to provide reference information for traffic management.The CS compression layer is located between the input layer and the first convolutional layer of the network.Its output measurement characteristic is used as the input of the convolution layer so that the amount of data to be processed by the entire classification network is greatly reduced,and the efficiency of the classification network is improved.Meanwhile,the measurement process of CS compression layer is obtained by the convolution of the image block and the filter formed from the sensing matrix component,and each measurement value contains key information of the original image,thereby the classification accuracy is improved.The experimental results show that our method is able to speed up the window-calibrating stages of CNN-based image classification.Moreover,our proposed method has better overall performance in vehicle-type detection compared with other methods.The method proposed in this thesis has higher accuracy in vehicle classification,and the classification accuracy is increased by 1.96%-5.87% compared with the existing methods,and the inference time is shortened by about 4-18 s.