Adaptive compressed sensing for wireless image sensor networks

Transcription

1 DOI /s x Adaptive compressed sensing for wireless image sensor networks Junguo Zhang 1 & Qiumin Xiang 1 & Yaguang Yin 2 & Chen Chen 3 & Xin Luo 1 Received: 30 April 2015 / Revised: 20 January 2016 / Accepted: 24 March 2016 # Springer Science+Business Media New York 2016 Abstract Compressed sensing (CS) based image compression can achieve a very low sampling rate, which is ideal for wireless sensor networks with respect to their energy consumption and data transmission. In this paper, an adaptive compressed sensing rate assignment algorithm that is based on the standard deviations of image blocks is proposed. Specifically, each image block is first assigned a fixed sampling rate. In addition to the fixed sampling rate, an adaptive sampling rate is then given to each block based on the standard deviation of the block. With this adaptive sampling strategy, higher sampling rates are assigned to blocks that are less compressible (e.g., blocks with complex textures are less compressible than blocks with a smooth background). The sensing matrix is constructed based on the assigned sampling rate. The fixed measurements and the adaptive measurements are concatenated to form the final measurements. Finally, the measurements are used to reconstruct the image on the decoding side. The experimental results demonstrate that the proposed algorithm can achieve image progressive transmission and improve the reconstruction quality of the images. * Junguo Zhang zhangjunguo@bjfu.edu.cn Qiumin Xiang xiangqiumin16@126.com Yaguang Yin yinyaguang@abs.ac.cn Chen Chen chenchen870713@gmail.com Xin Luo luoxinstudy@126.com School of Technology, Beijing Forestry University, Beijing , China Academy of Broadcasting Science, SAPPRFT, Beijing, China Department of Electrical Engineering, University of Texas at Dallas, Richardson, TX 75080, USA

2 Keywords Wireless sensor networks. Block compressed sensing. Adaptive sampling. Rate allocation 1 Introduction In recent years, wireless sensor networks have increasingly been applied to various fields, such as surveillance, tracking, disaster monitoring, home automation, industrial control, and battlefield surveillance, because of their low cost and ease of use [1, 13, 20, 21]. However, WSNs cannot effectively address the large amount of image data that is involved in using image sensors because the bandwidth, computing capability of the mobile processor and energy (e.g., battery) are limited. However, compressed sensing (CS) [2, 3, 7] has provided a new opportunity for solving this problem. Under the conditions of CS theory, a signal can be recovered from far fewer samples than the number required by the Nyquist sampling theorem, which reduces the amount of image data and computational complexity. In the applications that are based on WSNs, many image and video processing applications have been developed based on WSNs, for example, image compression and reconstruction, image fusion, image quality prediction, dynamic scene reconstruction [17, 18], and video coding, transmission and reconstruction [10, 15, 16]. However, the normal work of the WSNs requires not only highperformance mobile processors to process images but also a large volume of storage space to store the images. Compared to scalar data (such as temperature, soil moisture), a large amount of image data creates a burden for image processing and transmission and also poses a great challenge for WSNs due to the characteristics of restricted resources. Therefore, there is a pressing need for automated algorithms to preprocess and selectively compress data from raw images, reduce the amount of data, and save energy and storage space, thereby facilitating and supporting WSN applications. Traditional image compression algorithms based on the Nyquist sampling theorem (e.g., JPEG and JPEG2000) are not applicable for resource-restricted WSNs because of their high coding complexity and channel errors [8]. Recently, CS theory has broken the limit of the Nyquist sampling theorem and now provides a new approach for image acquisition and compression. The fundamental idea behind CS is that sampling and compressing are synchronous instead of two independent processes. Thus, CS can effectively reduce the amount of data of the image processing procedure. Even if some measurements are lost during transmission, the quality of the reconstructed signal will not be seriously affected, which leads to better robustness than traditional compression algorithms. These merits of CS motivate us to develop an image compression algorithm that is based on CS for image compression and transmission in WSNs. In a typical CS system for image processing, each measurement is the projection of an image in vector form onto an individual random vector. When the size of the signal is large, an enormous amount of memory is required to store the random sampling operator. Therefore, classical CS is not directly suitable for large-scale applications, such as high density signals and high resolution images. In [6, 9, 12], the idea of block compressive sensing (BCS) was proposed to reduce the size of the sensing matrix, in which the whole image was divided into small blocks and image sensing and reconstruction were conducted in a block-by-block manner. Therefore, the computational complexity of sampling and reconstruction are greatly reduced. However, the assignment of the sampling rate is fixed in such a way that the sampling rate of each block is identical without considering the structures of the blocks. To overcome this drawback of the uniform sampling strategy, several adaptive sampling rate assignment schemes have been

3 developed for CS image reconstruction. An adaptive wavelet packet image CS algorithm was proposed in [11]. First, the algorithm decomposes an image by a wavelet packet transform and then uses the mathematical expectation and information entropy as criteria to classify the wavelet packet transform coefficients of each block. Although this algorithm improves the quality of the reconstructed image, it does not use the BCS scheme. More importantly, the complexity of the algorithm is very high, and thus, it cannot be applied to WSNs. In [19], the sparsity of each image block in the DCT domain was chosen as the criterion to obtain measurements adaptively for each image block. However, this criterion uses a fixed threshold to allocate the sampling rates, which does not achieve a real adaptive sampling rate allocation. In [14], an adaptive image BCS algorithm that is based on texture information was proposed. To achieve adaptive sampling rate allocation, the algorithm must reconstruct the image first in the sampling process, which increases the computational complexity at the encoding side. An image CS algorithm that is based on an adaptive learning sparse matrix was proposed in [22]. Although the algorithm achieves high image reconstruction quality, the formation of its sparse groups requires extensive training, which limits its practical deployment. The aforementioned algorithms verified adaptive sampling strategies can improve the image reconstruction quality, but they are all sacrificed to increase the computational complexity at the encoding side, therefore preventing their use in energy-constrained WSNs. Considering this situation, we propose an adaptive CS algorithm for image transmission in WSNs. This algorithm overcomes the excessive size of the measure matrix, which is caused by very large input images as well as the loss of the image texture information due to the fixed sampling strategy. The algorithm uses an adaptive sampling rate allocation strategy that is based on the standard deviations of image blocks to improve the quality of the reconstructed images. Because WSNs have a low transmission bandwidth and few unstable channels, the fixed measurements and adaptive measurements are transmitted independently to achieve the progressive transmission of images using the proposed adaptive rate allocation strategy. The remainder of this paper is organized as follows. Section 2 presents the proposed block adaptive CS algorithm based on the standard deviations of image blocks. Section 3 provides the experimental results and analysis. Finally, Section 4 concludes the paper. 2 The proposed adaptive compressed sensing algorithm 2.1 Overall framework of the adaptive compressed sensing method In BCS, each CS measurement contains global information for an image block. The number of measurements that contain useful information has a large impact on the reconstruction result. The original BCS algorithm assigns the same sampling rate for each image block. However, when an image is divided into small blocks, different image blocks can contain different amounts of information. Blocks that contain little information can be reconstructed by fewer observations than those that contain more information. Thus, it is possible to adaptively allocate different sampling rates for different blocks to improve the BCS reconstruction performance. The overall framework of our proposed adaptive CS algorithm for images is presented in Fig. 1. This framework includes the adaptive sampling rate assignment, CS sampling, quantization, coding and CS reconstruction. In this paper, we focus on the adaptive sampling assignment for different blocks.

4 x1 x2 x3 xi Fixed measurement F matrix y F Fixed measurements F F F F y1 y y 2 3 y i Quantization Coding Block Adaptive measurement A matrix Original image Adaptive sampling rate allocation based on standard deviation A A A A y y y i A Adaptive measurements Information of matrix structure Measurement y1 y2 y3 matrix y y yi Channel A y SPL Inverse quantization Decoding F y Reconstructed image Measurements A F y [ y, y ] Fig. 1 Overall architecture of the proposed adaptive CS algorithm based on the standard deviations of the image blocks 2.2 Adaptive sampling rate allocation Standard deviation is often used to reflect the dispersion degree of a dataset in the science of probability theory and mathematical statistics. The concrete implementation of the adaptive sampling rate allocation method based on the standard deviation is shown in Fig. 2. Suppose that there is an image block in vector form x =(x 1, x 2,, x n ) T,theaverageofthe image pixels is x, and the standard deviation is std(x). If the value of the standard deviation is large, then there is a large difference between the value of each pixel and the average value in an image block. If the value of the standard deviation is small, then most of the pixels are relatively close to the average value of the pixels in a block. It is reasonable to consider that a non-smooth block contains more texture information than a smooth block. The difference between a pixel value and the average pixel value is large in a non-smooth block. Therefore, a high sampling rate is required to reconstruct the detailed information in a non-smooth block. On the other hand, a smooth block contains less texture information, and thus, a low sampling rate could be sufficient to reconstruct the information. Thus, the compressibility of an image block x i can be estimated by std(x i ), and the standard deviation can be used as the adaptive sampling rate allocation criterion. On this basis, we propose an adaptive sampling rate allocation strategy that is based on the standard deviations of image blocks. Step 1 Given the sampling rate SR, the number of image blocks n, and the block size (B B) of an image, the total sampling frequency M can be obtained by M = SR B 2 n. To ensure that the sampling frequency is less than the total number of pixels in an image block, we set the upper bound of the adaptive sampling frequency to upper =0.4 B 2 in this paper. Step 2 To guarantee the basic quality of the reconstructed image, we first give the same fixed sampling rate FSR = W SR to each image block, where W (0 W 1) is the fixed

5 Fig. 2 Flow chart of the adaptive sampling rate assignment strategy based on the standard deviations of image blocks sampling rate distribution parameter. If W is larger, then the fixed sampling rate of the image block is higher. Step 3 According to the fixed sampling rate FSR, the fixed sampling frequency of each image block can be obtained using FM i ¼ FSR M n. Step 4 Divide the original image into blocks and calculate the standard deviation of each image block std(x i ) and the percentage of each block standard deviation according to P i ¼ std ð x iþ n. stdðx i Þ i¼1 Step 5 Calculate the adaptive sampling frequency for each block: AM i = P i (M nfm i ). Step 6 Step 7 Step 8 If the adaptive sampling frequency of a block is greater than the upper bound, then compute the sum of the excessive part S = S+(AM i upper) and assign S to all of the blocks on average. After the rate assignment is completed, the adaptive sampling frequency can be beyond the upper bound again. Repeat step 6 until all of the blocks have their sampling rate within the range. Then, the adaptive sampling frequency of each block AM i can be obtained. Obtain the final fixed sampling frequency FM i and the adaptive sampling frequency AM i of each image block.

6 The block partition result is shown in Fig. 3(a), and the adaptive sampling rate of each image block allocation is shown in Fig. 3(b). As can be seen from this figure, the image blocks contain rich texture information (e.g., edges and corners) are given higher sampling rates than those contain less texture information (a) Adaptive sampling rate allocation for each block Without adaptive sampling rate allocation 0.7 rate sampling image blocks Fig. 3 a Block partition of an image (Lena). b Adaptive sampling rate allocation for each block in the image of Lena (b)

7 Table 1 The relationship between the PSNR (db) of the reconstructed image and the block size Algorithm PSNR PSNR PSNR PSNR STD-BCS-SPL Adaptive CS algorithm for the images based on the standard deviation can transmit fixed measurements and adaptive measurements to complete the progressive transmission and is suitable for WSNs with a low transmission bandwidth and unstable channel. For image reconstruction using the received measurements, we use the smoothed projected Landweber (SPL) algorithm developed in [12] due to its efficiency. Therefore, our proposed adaptive CS algorithm is called STD-BCS-SPL as an abbreviation. In [6], a post-processing that uses multihypothesis prediction [4, 5] to refine the reconstruction result of the BCS-SPL algorithm was developed. We could also utilize this strategy if there is no resource restriction on the decoder side. According to the specific needs, in some cases, we can choose to reconstruct the image only by fixed measurements to obtain the image information in a timely manner, and then, the fixed measurements are combined with adaptive measurements to improve the image quality. 3 Experiments and results analysis 3.1 Parameters selection The block size B is an important parameter in BCS. If B is too small, then there could be an increase in the computational complexity of computing the standard deviations of the blocks in our method. If B is too large, then the standard deviation of an image block might not accurately reflect the compressibility of a block. The fixed sampling rate allocation parameter W guarantees a basic quality of the reconstructed image and affects the adaptive sampling rate. Therefore, we must identify the optimal values for the two main parameters (B and W) ofthe STD-BCS-SPL algorithm. To tune these two parameters, we use the standard image BBarbara^ for this experiment. Table 1 shows the relationship between the peak signal-to-noise ratio (PSNR) value and the block size of an image when the sampling rate is 0.5 and the parameter W is 0.5. As can be observed from the Table 1, as the block size increases, the PSNR value gradually reduces. This relationship indicates that the standard deviation might not characterize the compressibility of an image block when the block size is large. Table 2 shows the relationship between the measurement sampling time (the sampling rate assignment and measurement sampling) and the block size of an image when the sampling rate is 0.5 and the parameter W is 0.5. From the results of this table, the sampling time of the STD-BCS-SPL algorithm is the shortest when the Table 2 The relationship between the sampling time (s) and the block size Algorithm Time Time Time Time STD-BCS-SPL

8 Fig. 4 The relationship between PSNR (db) of the reconstructed image and parameter W PSNR W block size is 8 8. Therefore, the block size has a significant influence on both the sampling speed and the quality of the reconstructed images. In this paper, we set the block size B to 8 in terms of the sampling speed and reconstruction performance. Lena Barbara Goldhill Barbara2 Boat Cameraman Fig. 5 Test images in our experiments

9 Figure 4 shows the relationship between the PSNR of the reconstructed image and the fixed sampling rate allocation parameter W when the overall sampling rate is 0.5 and the block size is 8 8. As illustrated in the figure, when W increases, the PSNR of the reconstructed image gradually decreases. This relationship verifies that using the adaptive sampling rate allocation (i.e., using a small fixed sampling rate (small W)) could improve the quality of the reconstructed image. The parameter W is set to 0.5 in our experiments. 3.2 Comparison experiment on the algorithm s performance In the experiment, the performance of the proposed algorithm (STD-BCS-SPL) is compared with the original BCS-SPL algorithm [12]. Although we review several adaptive sampling rate assignment schemes in the introduction, they are not suitable or practical for WSNs due to their high computational complexity in the encoder side. Therefore, in the context of resourcerestricted WSNs applications, the original BCS-SPL algorithm using a uniform sampling strategy can be considered as the state-of-the-art CS reconstruction algorithm for image compression and transmission in WSNs. To evaluate the reconstructed image quality, we use the PSNR as the image quality measurement. Six standard images are used in our experiments, including Lena, Barbara, Goldhill, Barbara2, Boat and Cameraman, as shown in Fig. 5. All of the images have the size of pixels. The experiments are conducted using MATLAB v7.8 (R2009a) on a PC with an Intel(R) Core(TM) i CPU at 3.20GHz and 4-GB of RAM. The sparse transform is Table 3 PSNR (db) comparison of the reconstructed images Algorithm Sampling rate PSNR PSNR PSNR PSNR PSNR Lena BCS-SPL STD-BCS-SPL Barbara BCS-SPL STD-BCS-SPL Goldhill BCS-SPL STD-BCS-SPL Barbara2 BCS-SPL STD-BCS-SPL Boat BCS-SPL STD-BCS-SPL Cameraman BCS-SPL STD-BCS-SPL

10 the wavelet transform, and the measurement matrix is a random Gaussian matrix. It should be noted that all of the reconstruction quantity evaluations (PSNR) are averaged over ten independent trials because the performance of the reconstruction varies due to the randomness of the sampling matrix. Table 3 shows the results of the two algorithms. Compared with the original BCS-SPL algorithm, which uses a uniform sampling strategy, the PSNR of the reconstructed image using the proposed algorithm is improved in the experiments (sampling rate = 0.1 is an exception). Particularly when the sampling rate is higher, the performance improvement is more obvious. This finding occurs because the increase in the total number of samples causes the allocation of higher sampling rates for image blocks that have detailed textures, which results in the reconstructed image having better quality. The average PSNR gain of Lena is db, and the maximal gain is 2.49 db; the average PSNR gain of Barbara is db, and the maximal gain is 3.42 db; the average PSNR gain of Barbara2 is 1 db, and the maximal gain is 2.36 db; and the average PSNR gain of Cameraman is db, and the maximal gain is 7.26 db, which is a remarkable improvement. As shown in Table 1, when the sampling rate is 0.1, the performance of the proposed algorithm is worse than that of the original algorithm for some of the images. The analysis of the experimental results shows that the effect of the adaptive algorithm is affected by the roedeer roedeer2 reddeer roedeer3 Fig. 6 Test wild animal images captured by wireless image sensor network for wildlife monitoring

11 sampling number of each image block. At a very low overall sampling rate (e.g., sampling rate = 0.1), the total sample number is small, and the adaptive allocation has little effect on the sampling number of each image block. It should be noted that it is reasonable to have the overall sampling rate in the range of 0.3 to 0.5 in order to have visually pleasing reconstructed images for practical applications, for example, wildlife monitoring using WSNs. We also evaluate our proposed method using real world images from wireless image sensor network. The images used in this experiment were collected from Saihanwula Nature Reserve (Mongolia, China) between May 2012 and August Specifically, the images were captured by an LTL 5210 MQ professional camera which was utilized as an image sensor node in the wireless image sensor network for wildlife monitoring. The four wild animal images were shown in Fig. 6. The reconstruction results of BCS-SPL and our proposed STD-BCS-SPL algorithms for the 4 real world wild animal images are reported in Table 4. The average PSNR gain of roedeer is db, and the maximal gain is 2.93 db; the average PSNR gain of roedeer2 is db, and the maximal gain is 5.14 db; the average PSNR gain of reddeer is db, and the maximal gain is 1.42 db; the average PSNR gain of roedeer3 is db, and the maximal gain is 0.51 db. Therefore, the experimental results on real world images captured by wireless image sensor network also demonstrate the superior performance of our proposed STD-BCS- SPL algorithm over the original BCS-SPL method. Finally, we show some visual comparisons between the proposed STD-BCS-SPL and the original BCS-SPL in Fig. 7. As observed from the Fig. 7, the reconstructed images using the proposed algorithm show more detail (e.g., texture) than the reconstructed images using the BCS-SPL algorithm. This finding arises because in STD-BCS-SPL relatively high sampling rates are assigned to the edge and rich texture blocks and low sampling rates are assigned to the smooth blocks. Therefore, fine details can be well preserved in the reconstructed images using the proposed STD-BCS-SPL algorithm. Compared to the uniform sampling strategy in the BCS-SPL, our method has more freedom on the sampling rates for different image blocks. Table 4 PSNR (db) comparison of the reconstructed images Algorithm Sampling rate PSNR PSNR PSNR PSNR PSNR roedeer BCS-SPL STD-BCS-SPL roedeer 2 BCS-SPL STD-BCS-SPL reddeer BCS-SPL STD-BCS-SPL roedeer3 BCS-SPL STD-BCS-SPL

12 Original image (cameraman) Magnified part (red box) of the original image BCS-SPL Proposed STD-BCS-SPL Original image (roedeer3) Magnified part (red box) of the original image BCS-SPL Proposed STD-BCS-SPL Fig. 7 Local magnification of the reconstructed images (cameraman and roedeer3) by different algorithms at sampling rate = 0.5

13 4 Conclusions The image CS algorithm is suitable for the Bsimple coding and complex decoding^ characteristics of WSNs and is one solution that provides efficient image transmission in WSNs with limited resources. In this paper, based on the BCS-SPL algorithm, we proposed an adaptive CS algorithm that is based on the standard deviation (STD-BCS-SPL). The proposed algorithm utilizes an adaptive sampling strategy that is based on the standard deviations of image blocks to optimize the sampling resources. Combined with the sampling rate allocation strategy, the fixed measurements and adaptive measurements are transmitted separately to achieve progressive transmission of the image. Compared with the original BCS-SPL algorithm, the experimental results show that the algorithm can effectively improve the quality of the reconstructed images. Moreover, the progressive transmission of the images is suitable for WSNs that have low transmission bandwidth and few unstable channels. Acknowledgments This work was funded by The National Natural Science Foundation of China (Grant No ), Import Project under China State Forestry Administration (Grant No ), Beijing Higher Education Young Elite Teacher Project (Grant No.YETP0760). References 1. Akyildiz I, Su W, Sankarasubramaniam Y, Cayirci E (2002) Wireless sensor networks: a survey. Comput Netw 38(4): Candès EJ (2006) Compressive sampling. Marta Sanz Solé 17(2): Candès EJ, Romberg J, Tao T (2006) Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency information. IEEE Trans Inf Theory 52(2): Chen C, Fowler JE (2012) Single-image super-resolution using multihypothesis prediction. Signals, Systems and Computers (ASILOMAR), 2012 Conference Record of the Forty Sixth Asilomar Conference on. IEEE 5. Chen C, Li W, Tramel EW, Cui M, Prasad S, Fowler JE (2014) Spectral-spatial preprocessing using multihypothesis prediction for noise-robust hyperspectral image classification. IEEE J Sel Top Appl Earth Obs Remote Sens 7(4): Chen C, Tramel EW, Fowler JE (2011) Compressed-sensing recovery of images and video using multihypothesis predictions. In: Proceedings of the 45th Asilomar Conference on Signals, Systems, and Computers, Pacific Grove, CA, pp Donoho DL (2006) Compressed sensing. IEEE Trans Inf Theory 52(4): Ferrigno L, Marano S, Paciello V, Pietrosanto A (2005) Balancing computational and transmission power consumption in Wireless Image Sensor Networks. In: Proceedings of the 2005 I.E. International Conference on Virtual Environments, Human-Computer Interfaces and Measurement Systems, Italy, pp Gan L (2007) Block compressed sensing of natural images. In: Proceedings of th International Conference on Digital Signal Processing, pp Liu Q, Yang Y, R-R J, Gao Y, Yu L (2012) Cross-view down/up-sampling method for multiview depth video coding. IEEE Signal Process Lett 19(5): Luo M-R, Zhou S-W (2013) Adaptive wavelet packet image compressed sensing. J Electron Inf Technol 35(10): Mun S, Fowler JE (2009) Block compressed sensing of images using directional transforms. In: Proceedings oficip,cairo,egypt,pp Tavli B, Bicakci K, Zilan R et al (2012) A survey of visual sensor network platforms. Multimedia Tools Appl 60(3):

14 14. Wang R-F, Jiao L-C, Liu F, Yang S-Y (2013) Block-based adaptive compressed sensing of image using texture information. Acta Electron Sin 41(8): Yang Y, Deng H-P, Wu J, Yu L (2015) Depth map reconstruction and rectification through coding parameters for mobile 3D video system. Neurocomputing 151(2): Yang Y, Liu Q, Liu H, Yu L, Wang F-L (2015) Dense depth image synthesis via energy minimization for three-dimensional video. Signal Process 112: Yang Y, Wang X, Guan T, Shen J-L, Yu L (2014) A multi-dimensional image quality prediction model for user-generated images in social networks. Inf Sci 281: Yang Y, Wang X, Liu Q, Xu M-L, Yu L (2015) A bundled-optimization model of multiview dense depth map synthesis for dynamic scene reconstruction. Inf Sci 320: Zhang S-F, Li K, Xu J-T, Qu G-C (2012) Image adaptive coding algorithm based on compressive sensing. J TianjinUniv45(4): Zhang J-G, Li W-B, Zhao X-L, Bai X-D, Chen C (2009) Simulation and research on data fusion algorithm of the wireless sensor network based on NS2. In: Proceedings of 2009 WRI World Congress on Computer Science and Information Engineering, Los Angeles, CA, March 7, pp Zhang J-G, Luo X, Chen C, Liu Z, Cao S (2014) A wildlife monitoring system based on wireless image sensor networks. Sens Transducers 180(10): Zhang J, Zhao C, Zhao D-B, Gao W (2014) Image compressive sensing recovery using adaptively learned sparsifying basis via L0 minimization. Signal Process 103: Junguo Zhang received his B.S. and M.S. degrees in China University of Mining and Technology, in 2000 and 2003, respectively, and the D.E. degree in Beijing Forestry University. He visited the Forest Product Laboratory, USDA in He is the director of the department of automation now. He is committed in the research on the forestry information collection and intelligent processing. In addition, he has led nearly ten scientific projects supported by the National Natural Science Foundation of China, State Forestry Administration, etc.

15 Qiumin Xiang is currently working toward the B.E. degree in automation from Beijing Forestry University, Beijing, China. His research interests include compressed sensing, signal and image processing. Yaguang Yin received his B.S., M.S. and Ph.D. in Department of Automation, Tsinghua University, China in 2000, 2003 and He is now a Research Fellow in Academy of Broadcasting Science, State Administration of Press, Publication, Radio, Film and Television of China. His research interests include image and video coding, vision computation in big data, etc.

16 Chen Chen received the B.E. degree in automation from Beijing Forestry University, Beijing, China, in 2009 and the M.S. degree in electrical engineering from Mississippi State University, Starkville, MS, in He is currently working toward the Ph.D. degree in the Department of Electrical Engineering, University of Texas at Dallas, Richardson, TX. His research interests include compressed sensing, signal and image processing, hyperspectral image analysis, pattern recognition, and computer vision. He is an active reviewer for the IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, IEEE TRANSACTIONS ON IMAGE PROCESSING, and IEEE SIGNAL PROCESSING LETTERS. Xin Luo received the B.E. degree in automation from Beijing Forestry University, Beijing, China, in He is currently working toward the M.S. degree in School of Technology from Beijing Forestry University, Beijing, China. His research interests include wireless sensor networks, compressed sensing, signal and image processing.