In modern agriculture, it is essential to monitor plant propagation or growth for guaranteeing productivity. The labor costs can be efficiently reduced by automatically capturing the architectural parameters of the plant, so more and more attention has been paid to the design of the plant monitoring system (Somov et al., 2018; Grimblatt et al., 2021; Rayhana et al., 2021). Early, lots of systems are designed to monitor the various environmental parameters on plant growth, such as humidity, temperature, solar illuminance, etc., e.g., James and Maheshwar (2016) used multiple sensors to measure the soil data of plants and transmitted these data to the mobile phone by Raspberry Pi; Okayasu et al. (2017) developed a self-powered wireless monitoring device that is equipped with some environmental sensors; Guo et al. (2018) added big-data services to analyze the environmental data on plant growth. These environmental parameters indirectly indicate the process of plant growth, and they cannot record the visual scenes on plant growth, resulting in the unavailability of the physical structure parameters on plants. To realize the visual monitoring of plants, some works have started to integrate the visual sensors into the plant monitoring system, e.g., Peng et al. (2022) used the binocular camera to capture video sequences on a plant and used the structure from motion method (Piermattei et al., 2019) to extract the 3-D information of a plant; Sajith et al. (2019) designed a complex network to derive the plant growth parameters from the monitoring images; Akila et al. (2017) extracted the plant color and texture by the visual monitoring system. From the above, it can be seen that the visual sensor or camera is used to capture the video sequences on plant growth, and these video sequences are compressed as bitstream which is transmitted to the IoT cloud for further analyzing. As the core of visual sensors, the video compression is a major energy consumer, so a challenge that we face for the visual monitoring system of the plant is to design an energy-efficient video coding scheme to extend the working time of the visual sensor. In the framework of IoT, CVS is a potential coding scheme to reduce the energy consumption of visual sensors. The following briefly overviews the CVS systems.

Compressive Video Sensing is the marriage of CS theory and DVC, which reduces the encoding costs and enhances the robustness to noise, thus becoming a potential video codec for wireless visual sensors. At the encoder, to satisfy low complexity and fast computation, the block-based CS sampling is performed on each video frame independently, i.e., the ith video frame f_i of size N₁ × N₂ is partitioned into non-overlapping blocks of size B × B, each block is vectorized as x_i,j of length N_b, and the CS measurements y_i,j of x_i,j are output by (1)yi,j=Φi,j·xi,j where Φ_i,j is called as the measurement matrix and can be constructed by some random matrices, e.g., Gaussian, Bernoulli, structural random matrix, etc. By setting the length of y_i,j to be M_i,j, the size of Φ_i,j is fixed to be M_i,j × N_b, and the subrate S_i of f_i is defined as (2)Si=MiN=∑j=1JMi,jN1×N2 where N is the number of total pixels in f_i, M_i is the number of CS measurements for f_i, and J is the number of blocks in f_i. In CI application, an optical device is designed to perform Equation (1), and directly output the CS measurements. To ensure a stable recovery, L video frames are gathered to form a GOP, in which the first frame, called the key frame, is set to be a high subrate, and others, called the non-key frame, are set to be a low subrate. After quantization, all CS measurements of GOP are packaged and transmitted to decoder.

At the decoder, by using the received CS measurements, the frame-by-frame, 3D, or distributed strategy is performed to reconstruct the GOP. For frame-by-frame, the reconstruction model can be represented by (3){x^i,j}j=1J=argmin{xi,j}j=1J                {∑j=1J||yi,j-Φi,j·xi,j||22+α·∑j=1J||Ψ·xi,j||1} where Ψ denotes the 2D sparse representation basis, α is a regularization factor, ||·||₂ denotes ℓ₂ norm, and ||·||₁ denotes ℓ₁ norm. The model (3) can be solved by some non-linear optimization algorithms, e.g., Alternating Direction Method of Multipliers (ADMM) (Yang et al., 2020), and all reconstructed blocks are spliced into the estimated frame f^i. The frame-by-frame model uses only the spatial correlations, so its rate-distortion performance is unsatisfactory. The 3D reconstruction model fully considers the spatial-temporal correlations and it can be represented by (4){x^i,j|i=1L}j=1J=argmin{xi,j|i=1L}j=1J{∑i=1L∑j=1J∥yi,j-Φi,j·xi,j∥22                                         +α·∑i=1L∑j=1J|Γ·[x1,j,x2,j,⋯,xL,j]|} where Γ denotes the 3D sparse representation basis, and it is used to remove the spatial-temporal redundancies between blocks. Though effective, model (4) results in a heavy computational burden. Different from the 3D reconstruction, the distributed reconstruction uses the motion-compensation based prediction technique to expose the spatial-temporal redundancies between blocks. Figure 1 shows the mechanism of MH prediction, which is commonly used in distributed reconstruction. MH prediction collects the spatial-temporal neighboring blocks in key frames to construct an MH matrix H_i,j. According to the motion vector v_i of x_i,j, the motion-aligned windows W₁ and W₂ of sizes W × W are, respectively, located on the previous and the next key frames, and all candidate blocks in W₁ and W₂ are extracted as the hypotheses {ht}t=1T of x_i,j, producing H_i,j = [h₁, h₂, ⋯ , h_T], in which T = W². By using MH prediction, the distributed reconstruction is modeled as a Least-Squares (LS) problem as follows: (5)ŵi,j=argminw{∥yi,j-Φi,j·Hi,j·w∥22+β·∥Θ·w∥22} (6)x^i,j=Hi,j·ŵi,j where Θ is the Tikhonov matrix, and β is a regularization factor. Θ is a diagonal matrix and constructed by (7)Θ=[∥yi,j-Φi,j·h1∥20⋱0∥yi,j-Φi,j·hT∥2] With this structure, Θ assigns weights of small magnitude to hypotheses mostly dissimilar from x_i,j. The LS problem can be fast solved by the Conjugate Gradient algorithm (Zhang et al., 2018), which significantly reduces the computational complexity of distributed reconstruction. Due to the full exploitation of spatial-temporal correlations between blocks, the MH prediction enables the distributed reconstruction to provide superior recovery. From the above, in order to realize a light decoder and ensure a good recovery at the same time, distributed reconstruction is a wise way.

Compressive Sensing theory indicates that the sparsity K of the signal determines its required number M of CS measurements by precise recovery. An empirical rule (Becker and Bobin, 2011) is that the precise recovery can be achieved if (8)M≥4·K In the block-based CS sampling, this rule can be used to avoid the redundancy or insufficiency of CS measurements for blocks, i.e., adapted by the sparsity, each block is allocated to the appropriate number of CS measurements. The sparsity is defined as the number of coefficients with significant magnitude in a representation, and its calculation has not a strict mathematical formula. For images, the sparsity can be revealed by some features, e.g., edge, variance, gradient, etc., and these features are applied into adaptive allocation, leading to the improvement of recovery quality. The simple features only describe the correlations between pixels, but the structures of blocks are not taken into consideration, thus we require some complex features to improve the efficiency of adaptive allocation. In Ref. Romano and Elad (2016), the self-similarity descriptor (Shechtman and Irani, 2007) is used to extract the contexts of blocks, which represents how similar a central block is to its large surrounding windows. Contexts contain the internal structures and external relations among blocks, and it is a potential feature to better reveal the sparsity variation. The following briefly describes how to extract the contexts in an image.

The context feature expresses the similarities between a central block and those of its large surrounding windows. As illustrated in Figure 2, for a central block x_p in an image, its similarity weights are computed by (9)sp,q=exp{-∥xp-xq∥222σ2},∀q∈Ωd(p) where x_q denotes the qth surrounding block in a neighborhood Ω_d(p) of size d × d, and σ is a normalization factor. The range of s_p,q is [0, 1], in which a large value indicates that the blocks x_p and x_q are highly similar, and a small value indicates that the two are substantially different. All weights constitute a correlation surface U_p = [s_p,q|∀q ∈ Ω_d(p)], of which the statistics reveal the self-similarity of x_p. To measure the statistics, the correlation surface of x_p is rearranged into a histogram of b bins, of which the normalization is regarded as the context feature g_p of x_p.

The context feature g_p is an empirical distribution of the co-occurrences of x_p in its large surroundings, which measures the correlations between x_p to its surroundings. When g_p is biased toward the left bins, it can be concluded that the majority of s_p,q are small, indicating the block x_p is unique, i.e., it originates from a highly textured and non-repetitive area, so its sparsity is relatively high. When g_p is biased toward the right bins, it means that most of s_p,q are high, indicating that the block x_p has many co-occurrences in its surroundings, i.e., it originates from a large flat area, so its sparsity is low. From the above, we can see that the context feature accurately describes the geometric structure of a block with respect to its surrounding blocks, thus it is naturally sensitive to the sparsity variation. However, in CVS, the traditional method is impractical due to the unavailability of original pixels or high computational complexity. Therefore, it is challenging to extract the context feature by using CS measurements of blocks.

As shown in Figure 3, we describe the architecture of the proposed context-based CVS system in detail. The input video sequence is divided into several GOPs of length L, and each GOP_k is successively encoded as Packet_k. After receiving this packet, the decoder reconstructs the corresponding GOP^k, and all reconstructed GOPs are regrouped as the entire video sequence.

Figure 3A presents the process of encoding GOP_k. The key frame f₁ is split from GOP_k, and others {fi}i=2L are regarded as the non-key frames. The key frame f₁ and the ith non-key frame f_i are partitioned into J non-overlapping blocks {x1,j}j=1J and {xi,j}j=1J of size B × B, respectively. For the key frame f₁, we set a high subrate S₁ = S_K to sample the blocks {x1,j}j=1J and generate the CS measurements {y1,j}j=1J according to Equation (1). The blocks {xi,j}j=1J in the non-key frame f_i are sampled at a low subrate S_i = S_NK, producing the corresponding CS measurements {yi,j}j=1J by Equation (1). For f₁ and f_i, based on the preset subrates, CS measurements are uniformly allocated to each block, however, without considering the structures of blocks, the uniform allocation results in either redundancy or insufficiency of CS measurements for some blocks. To improve the efficiency of block-based CS sampling, the core of the encoder is to perform the adaptive allocation by contexts of blocks. Different from traditional methods, the contexts U_1,j and U_i,j of x_1,j and x_i,j are, respectively, extracted by using the CS measurements y_1,j and y_i,j, which makes CVS system more practical. After context extraction, according to the contexts U_1,j and U_i,j, the numbers of CS measurements of x_1,j and x_i,j are modified as M_1,j and M_i,j by adaptive allocation. According to M_1,j and M_i,j, by removing the redundancy or supplementing the insufficiency in y_1,j and y_i,j, x_1,j and x_i,j are re-sampled as y~1,j and y~i,j, respectively. DPCM cannot be used to quantize the adaptive measurements with different numbers. To overcome this defect of DPCM, we fuse zero padding into DPCM and predictively quantize y~1,j and y~i,j as y~1,jq and y~i,jq. Finally, all quantized CS measurements are encoded as bits by Huffman and packaged as Packet_k.

Figure 3B presents the process of decoding Packet_k. After unpackaging Packet_k, the inversions of Huffman and zero-padding DPCM are implemented, and the CS measurements of x_1,j and x_i,j are recovered as y^1,j and y^i,j which have some quantization errors with their originals y~1,j and y~i,j. The distributed reconstruction is performed to reconstruct the key frame f₁ and the non-key frames {fi}i=2L. To suppress the blocking artifacts in the reconstructed frames, we realize the recovery of large blocks by merging the CS measurements of the spatially neighboring blocks, so the CS measurements of f₁ and f_i are updated as z_1,r and z_i,r for large blocks. Based on z_1,r, the reconstructed key frame f^1 is produced by using a linear recovery model, which rapidly recovers each block by a matrix-vector product. Regarding the previous and the next reconstructed key frames as references, the MH prediction outputs the reconstructed non-key frame f^i by using z_i,r. Finally, all reconstructed frames are combined into GOP^k. Details of the core parts, including contexts extraction, measurements allocation, zero-padding DPCM, and distribution reconstruction, are described in the following subsections.

In the proposed CVS system, the context features are extracted by using the CS measurements of blocks. As illustrated in Figure 4, we compute the correlation surface U_i,j of x_i,j in f_i as its contexts, in which i = 1, 2, ⋯ , L. In the surrounding window of size d_b × d_b centered on x_i,j, we cannot extract the original blocks pixel-by-pixel due to the unavailability of original pixels, but can only use the CS measurements {yi,j↺n}n=1Nc of non-overlapping blocks {xi,j↺n}n=1Nc, in which Nc=db2. According to CS theory, the measurement matrix Φ_i,j holds the Restricted Isometry Property (RIP) (Candès and Wakin, 2008) for blocks {xi,j}j=1J, which implies that all pairwise distances between original blocks can be well preserved in the measurement space, i.e.,

where it is noted that all blocks share the same measurement matrix Φ_i,j due to the uniform allocation. Based on Equation (10), the similarity weights between x_i,j and x_{i, j↺n} can be estimated by (11)si,j↺n=exp{-∥yi,j-yi,j↺n∥222σ2},∀n∈{1,2,⋯,Nc} All weights constitute the correlation surface U_i,j as follows: (12)Ui,j=[si,j↺n|∀n∈{1,2,⋯,Nc}] To compactly represent the contexts of x_i,j, we compute the mean u_i,j of U_i,j as the context feature, i.e.,

By exploiting the context feature u_i,j of x_i,j, we set the appropriate number of CS measurements for x_i,j, and remove the redundancy or supplement the insufficiency in y_i,j. The magnitudes of context features are high in smooth regions, and the magnitudes are low in the edge and texture regions, so it is found that the experience that the context feature is inversely proportional to the sparsity. Based on this experience, we can describe the distribution on the sparsity degrees of blocks by (14)Pi,j=ui,j-1∑j=1Jui,j-1 According to the present subrate S_i of f_i, we construct the allocation model of CS measurements for blocks as follows: (15)Mi,j=arg minmi,j∑j=1J(mi,j-Pi,j·Si·N)s.t. ∑j=1jmi,j=Si·N,mi,j≤0.9·Nb,mi,j∈ℕ+ where N is the total number of pixels in f_i, N_b is the block length, m_i,j is a positive integer, and its upper bound is set to be 0.9·N_b. The model (15) is solved according to Algorithm 1 and outputs the final number M_i,j of CS measurements for x_i,j.

Due to the adaptive allocation, the lengths of the re-sampled CS measurements {y~i,j}j=1L vary. Compared with SQ, DPCM provides better rate-distortion performance by adding the predictive scheme into the quantization of block-based CS measurements. However, DPCM requires that all blocks have the same number of CS measurements, as a result, DPCM cannot be used to quantize {y~i,j}j=1L. To make DPCM adapt to the adaptive allocation, we propose zero-padding DPCM, whose implementation is shown in Figure 5. Before inputting y~i,j to DCPM, we fill zeros in the last of y~i,j to make its length the same as others. After obtaining the de-quantized CS measurements y^i,j, we delete the zeros in the last of y^i,j to recover its original length M_i,j. By zero padding, each measurement in y^i,j-1 can be used to predict the corresponding measurement in y^i,j, and especially when there is predictive measurement ŷ_{i, j−1}(m) of the m-th measurement ỹ_i,j(m), the residual yi,jd(m) can be significantly reduced due to the intrinsic spatial correlation between y~i,j and y~i,j-1. The rate-distortion curves of the reconstructed Foreman, Mobile, and Football sequences are presented when zero-padding DPCM and SQ are, respectively, used to quantize the adaptive CS measurements (shown in Supplementary Figure 1), in which the rate-distortion curve is measured in terms of the Peak Signal-to-Noise Ratio (PSNR) in dB and bitrate in bits per pixel (bpp), and the linear recovery algorithm presented in subsection 3.5 is used to recover each video frame. It can be seen that zero-padding DPCM presents competitive performance with SQ at low bitrates but as the bitrate increases, its improvement of performance over SQ is increasingly significant. From these results, we find that the efficiency of zero-padding DPCM relies on the correlation between block-based CS measurements. With insufficient measurements, the correlation is weakened by the filling of excessive zeros, causing the performance degradation, but when measurements are sufficient, a high correlation is maintained, so the performance improvement stands out. From the above, zero-padding DPCM is more suitable for adaptive measurements compared with SQ.

At decoder, the distributed strategy is performed to reconstruct the key frame f₁ and the non-key frames {fi}i=2L, in which f₁ is estimated by a linear recovery model, and f_i is produced by MH prediction. To highlight the complex structures by contexts, a small block size is more desired at the encoder. However, the small block size causes serious blocking artifacts due to the differences of neighboring blocks in recovery quality. To suppress the blocking artifacts, we merge the CS measurements {y^i,j}j=1J of the small blocks {xi,j}j=1J into those {zi,r}r=1R of the large blocks {x~i,r}r=1R and realize the sampling of small blocks and the recovery of large blocks. The size B_lev × B_lev of large block is set to be (16)Blev=2lev·B,lev=1,2,⋯ in which lev is a positive integer. The number R of large blocks is N/Blev2, and it is smaller than the number J of small blocks. Figure 6 illustrates the block merging when lev is set to be 1. The four neighboring blocks x_i,j, x_i,j+1, x_i,j+N1/B, x_i,j+1+N1/B are merged into a large block x~i,r, and their CS measurements y^i,j, y^i,j+1, y^i,j+N1/B, and y^i,j+1+N1/B are spliced into z_i,r in rows, i.e., (17)zi,r=[y^i,jy^i,j+1y^i,j+N1/By^i,j+1+N1/B]≈Λi,r·[xi,jxi,j+1xi,j+N1/Bxi,j+1+N1/B] (18)Λi,r=[Φi,j  0 Φi,j+1    Φi,j+N1/B 0  Φi,j+1+N1/B] in which Λ_i,r is the diagonal matrix composed of the block measurement matrices Φ_i,j, Φ_i,j+1, Φ_i,j+N1/B, and Φ_{i, j+1+N1/B}, N₁ is the total number of rows in f_i, and B is the block size of the small block. To make z_i,r, the CS measurements of x~i,r, we transform x~i,r as (19)[xi,jxi,j+1xi,j+N1/Bxi,j+1+N1/B]=I·x~i,r in which I is an elementary column transformation matrix. Plugging Equation (19) into Equation (17), we build the bridge between x~i,r and z_i,r by (20)zi,r≈Λi,r·I·x~i,r=Ai,r·x~i,r in which A_i,r = Λ_i,r·I. According to Equation (20), the large block x~i,r can be recovered by using z_i,r. When lev is set to be larger than 1, the block merging can be done in manner similar to the above.

After the block merging, we use {z1,r}r=1R to recover the key frame f₁. The block x~1,r of f₁ is linearly estimated by (21)x^1,r=P1,r·z1,r in which P_1,r is the transformation matrix produced by the following model: (22)P1,r=arg minP{E[∥x~1,r-P·z1,r∥22]} in which E[·] denotes the expectation function. The model (22) outputs the optimal transformation matrix to minimize the mean square error between x~1,r and its estimator x^1,r, and it can be solved by making the gradient of objective function equal to 0, producing (23)P1,r=E[x~1,rz1,rT]E-1[z1,rz1,rT] Plugging Equation (20) into Equation (23), we get (24)P1,r=Corxx·A1,rT(A1,r·Corxx·A1,rT)-1 (25)Corxx=E[x~1,rx~1,rT] in which Cor_xx is the auto-correlation matrix of x~1,r, and its element Cor_xx[m, n] is estimated as follows: (26)Corxx[m,n]=0.95δm,n in which δ_m,n is the Euclidean distance between two pixels x~1,r(m) and x~1,r(n) in x~1,r. When the subrate is set to be large, the linear recovery model can provide excellent visual quality while costing fewer computations.

In the proposed CVS system, in order to highlight the complex structures by contexts, we desire a small block size at encoder, but at decoder, a large block size is desired to suppress the blocking artifacts in the reconstructed video frames. We set a block-size pair (B, B_lev), in which B and B_lev are the block sizes for sampling and recovery, respectively, and evaluate the effects of different block-size pairs on the reconstruction qualities of key frames and non-key frames.

First, we select the first frames of Foreman, BlowingBubbles, and ParkScene sequences as the key frames, which are linearly recovered, and show their rate-distortion curves at different block-size pairs in Figure 7. For Foreman and BlowingBubbles with the low resolution, the block-size pair (4, 16) achieves higher PSNR values than others with low bitrates, but the rate-distortion curve for the block-size pair (2, 16) rapidly increases as the bitrate increases and significant PSNR gains are achieved when compared with other block-size pairs. These results indicate that the small blocks used in adaptive allocation and large blocks for linear recovery fit together well. For ParkScene with high resolution, when the block size B for sampling is set to be too small, e.g., B = 2, no block can contain sufficient structures, causing the rate-distortion performance to degenerate as the bitrate increases, but a suitable block size for sampling is set, e.g., B = 8, PSNR gains can be significantly improved.

Then, we select the second frames of Foreman, BlowingBubbles, and ParkScene sequences as the non-key frames, which are recovered by MH prediction based on the reconstructed previous and next key frames at the subrate 0.7, and show their rate-distortion curves at different block-size pairs in Figure 8. Similar to the results from key frames, for Foreman and BlowingBubbles, the better rate-distortion performance is achieved when the block-size pair is set to be (2, 16), and for ParkScene, in order to prevent the loss of structures, the block size for sampling is appropriately set to be 8.

Given the above, we can see that the bad effects resulting from the extraction of contexts can be suppressed by the block merging, therefore, the quality improvement from contexts-based allocation is further enhanced.

In the proposed CVS system, the contexts are extracted from CS measurements and used to adaptively allocate the CS measurements for blocks, leading to the improvement of reconstruction quality. To verify the validity of contexts from CS measurements on the quality improvement, we evaluate the effects of different allocation schemes on the rate-distortion performance of the proposed CVS system. The uniform allocation is used as a benchmark, and the adaptive allocation uses the contexts extracted from CS measurements and original pixels, respectively.

Figure 9 shows the rate-distortion curves of the reconstructed key frames when using different allocation schemes, in which the key frames are, respectively, taken from the first frames of Foreman, BlowingBubbles, and ParkScene sequences. It can be seen that adaptive allocation outperforms uniform allocation in PSNR values at any bitrate, indicating that contexts contribute to quality improvement. Importantly, the contexts from CS measurements are competitive with those from original pixels, and their performance gaps are very small, which means that CS measurements can better represent the contexts of blocks.

Figure 10 shows the rate-distortion curves of the reconstructed non-key frames when using different allocation schemes, in which the non-key frames are, respectively, taken from the second frames of Foreman, BlowingBubbles, and ParkScene sequences. It can be seen that the adaptive allocation is still effective for MH prediction, and it can significantly improve the rate-distortion performances when compared with uniform allocation. The contexts from CS measurements have similar efficiency of allocation to that of contexts from original pixels, which proves that the merits of adaptive allocation can still be maintained in the measurement domain.

The above results indicate that the contexts extracted by CS measurements prompt the adaptive allocation to improve the reconstruction quality of CVS system, which makes the proposed CVS system more suitable to the applications with limited resources.

We evaluate the performance of the proposed CVS system by comparing it with the two state-of-the-art CVS systems: SS-CVS (Li et al., 2020) and MH-RTIK (Chen C. et al., 2020). To make a fair comparison, we keep the parameter settings of SS-CVS and MH-RTIK in their original reports, some important details are repeated as follows:

SS-CVS: the system consists of one base layer and one enhancement layer; the block size is set to be 16; the length of GOP is 10; the subrate of key frame is set to be 0.9; the dimension of the subspace is 10; the number of subspaces is 50.

In addition, we employ SQ and Huffman in SS-CVS and MH-RTIK to compress the CS measurements. For the proposed CVS system, the block-size pair is set to be (2, 16) for CIF and QWVGA sequences and (8, 16) for 1080P sequences, the subrate S_K of key frame is set to be 0.7, the results under the GOP length L = 2 are compared with those of MH-RTIK, and the results under the GOP length L = 10 are compared with those of SS-CVS.

Table 1 lists the average PSNR values for the reconstructed video sequences by the proposed CVS system, SS-CVS, and MH-RTIK when the subrate S_NK of non-key frame varies from 0.1 to 0.5. Compared with MH-RTIK, the proposed CVS system achieves obvious PSNR gains at any subrate, e.g., the average PSNR gain is 2.824 dB for the Foreman sequence. Compared with SS-CVS, the proposed CVS system also presents higher PSNR values at any subrate, and especially for low subrates, PSNR gains are significant, e.g., when the subrate is 0.1, PSNR gains are 9.82, 13.20, and 19.78 dB for ParkScene, BlowingBubble, Foreman sequences, respectively. Figures 11, 12 show the rate-distortion curves for the proposed CVS system, MH-RTIK, and SS-CVS. Due to the implementation of zero-padding DPCM, the performance improvement of the proposed CVS system is further enhanced when compared with MH-RTIK and SS-CVS. By the objective evaluation of the reconstruction quality, it can be indicated that the proposed CVS system can significantly improve the qualities of the reconstructed video sequences.

Table 2 lists the average encoding time (s/frame) and decoding time (s/frame) on video sequences with different resolutions for the proposed CVS system, SS-CVS, and MH-RTIK. We compute the average execution time on the range [0.1, 0.5] of subrate S_NK for the proposed CVS system and compare it with that of MH-RTIK for CIF sequences. The encoding speed of the proposed CVS system is slowed down due to the contexts-based adaptive allocation, and its encoding time is 0.63 s per frame, larger than that of MH-RTIK. Assisted by the simple linear recovery, the proposed CVS system reduces the decoding complexity, and only costs 4.48 s to reconstruct a video frame, however, MH-RTIK requires 19.34 s per frame. Under the subrate S_NK = 0.6, the execution time of the proposed CVS algorithm is compared with that of SS-CVS for the CIF, QWVGA, and 1080P video sequences, respectively. Compared with SS-CVS, the proposed CVS system costs less encoding time, and the encoding time does not dramatically increase as the resolution increases, e.g., for 1080P sequence, the proposed CVS system only costs 1.83 s per frame, but SS-CVS costs 108.10 s. In SS-CVS, the subspace clustering and the basis derivation are implemented at the encoder, and they lead to more encoding costs than the adaptive allocation in the proposed CVS system. The proposed CVS system costs less decoding time than SS-CVS, and its decoding costs also grow more slowly when compared with SS-CVS, e.g., for 1080P sequence, the proposed CVS system costs 162.47 s per frame, and the SS-CVS costs 401.8 s. The heavy computational burdens for SS-CVS derive from the non-linear subspace learning, but the decoding complexity of the proposed CVS system is limited benefiting from the linear recovery and prediction. From the above, we can see that the proposed CVS system still keeps a low computational complexity while providing better rate-distortion performance.