A Novel Video Compression Algorithm Based on Wireless Sensor Network

The video transmission on the wireless channel faces such problems as narrow bandwidth, great bandwidth volatility and high error rate. This calls for a video compression coding standard, which is suitable for a low bandwidth environment and capable of maximizing the video quality at a given code stream.

The videos processed by the H.264 coding standard support video communication. So far, many improved rate control algorithms have been developed based on this standard. For instance, Andrew Armstrong, [1] proposed a method to optimize the initial QP, in which the QP value is determined through binary search. [2] created an enhanced rate and rate-distortion model, which solves the “chicken or the egg causality dilemma” in H.264 standard using QP estimation and update, and enhances the accuracy by adjusting the number of bits occupied by the header information at a low bit rate. [3] set up a rate-distortion model in ρ domain according to the statistical relationship between the proportion of non-zero values of discrete cosine transform (DCT) coefficients in the coefficient table and the code rate.

In addition, many scholars have proposed error concealment techniques. For example, [4] developed a zero-motion vector method for images with weak motion intensity, which makes up for the data loss at the decoder with the macroblock at the same position in the reference frame. [5] put forward the boundary matching algorithm (BMA) capable of preventing boundary problems like edge breakage and object deformation; By this method, the error is concealed by multiple motion vectors in different regions of the damaged macroblock instead of a single motion of the macroblock.

In light of the above, this paper probes into the video compression technology in the wireless transmission environment, aiming to achieve efficient video encoding and decoding and make the encoded code stream suitable for transmission in wireless sensor network (WSN) channels. Two key problems are tackled in this research: How to handle the compressed video stream at the encoder end to make the bandwidth suitable for wireless transmission without sacrificing video quality; how to decode the error-containing video sequence received at the decoder end to minimize the video quality degradation in wireless transmission.

As its name suggests, the rate control problem aims to control the bit rate outputted by the encoder through adjustment of coding parameters (e.g. frame rate, motion detection threshold and QP) according to the actual state information (e.g. code rate, delay and buffer) and the rate-distortion theory. The philosophy of rate control is presented in Figure 1. During the rate control, the rate controller controls the output code stream of the video encoder, while the buffer smooths the output bit rate. The code rate at the input of the buffer is varying, while that at the output of the buffer depends on the rate control algorithm.

1.png

Figure 1. Philosophy of rate control

The rate control in this paper attempts to ensure the match between the video stream and the available network bandwidth, thereby mitigating network congestion and smoothing the output stream of the encoder.

3.1 H.264 rate control algorithm

The H.264 rate control algorithm was developed on the basis of the basic unit, the fluid motion traffic model, and the mean absolute difference (MAD) linear prediction model. If a basic unit is divided into sub-units in the size of one frame, then the rate control only involves two layers, namely, the group of pictures (GOP) layer and the frame layer. The GOP layer rate control algorithm includes two steps: determining the target number of bits and initializing the QP.

Step 1. Determining the number of bits to be coded for the GOP layer

Before the encoding of the i-th GOP, the total number of bits to be allocated to the GOP can be calculated as:

$T_{r}\left(n_{i, 0}\right)=\frac{u\left(n_{i, 1}\right)}{F_{r}} \times N_{g o p}-B_{c}\left(n_{i-1, N_{g o p}}\right)$   (2)

where $N_{g o p}$ is the number of frames contained in the GOP. Since the channel has a time-varying bandwidth, T_r is continuously updated in units of frames according to the following equation:

$T_{r}\left(n_{i, j}\right)=T_{r}\left(n_{i, j-1}\right)+\frac{u\left(n_{i, j}\right)-u\left(n_{i, j-1}\right)}{F_{r}}\left(N_{g o p}-j\right)-b\left(n_{i, j-1}\right)$   (3)

If the code rate is fixed, then $u\left(n_{i, j}\right)=u\left(n_{i, j-1}\right)$ can be simplified to equation (4), that is, equation (2) specifies the number of bits to be allocated to a GOP under a fixed code rate.

$T_{r}\left(n_{i, j}\right)=T_{r}\left(n_{i, j-1}\right)-b\left(n_{i, j-1}\right)$    (4)

Step 2. Initializing the QP

The initial QP, denoted as QP₀, is estimated based on the available channel bandwidth and the GOP length. QP₀ is the QP used to encode the I frame and the first P frame of the first GOP in the frame sequence. In general, the QP₀ value should be negatively correlated with the channel bandwidth, and reduced by a unit every time the GOP length increases by 30. The value of QP₀ can be calculated as:

$Q P_{0}=\left\{\begin{array}{ll}{40} & {b p p \leq l_{1}} \\ {30} & {l_{1}<b p p<l_{2}} \\ {20} & {l_{2}<b p p \leq l_{3}} \\ {10} & {b p p>l_{3}}\end{array}\right.$  (5)

$b p p=u\left(n_{0,0}\right) / F_{r} N_{p i x e l}$   (6)

where bpp is the number of bits occupied by each pixel; $N_{p i x e l}$ is the total number of pixels per each frame of image.

Except for the first GOP, the first QP QP_st in other GOPs can be calculated as:

$Q P_{s t}=\frac{\operatorname{sum}_{P Q P}}{N_{p}}-\frac{8 \times \operatorname{Tr}\left(n_{i-1, N_{g o p}}\right)}{\operatorname{Tr}\left(n_{i, 0}\right)}-\min \left\{2, \frac{N_{g o p}}{15}\right\}$    (7)

where N_p is the number of all P frames in the previous GOP; $Sum_{pep}$  is the sum of the QPs of all P frames in the previous GOP.

3.2 Improvement of rate control algorithm

In the GOP layer rate control of the H.264 standard, the QP of the I frame and the first P frame of the first GOP in the frame sequence, denoted as QP₀, is configured in advance, while the initial QPs of the other GOPs, denoted as QP_st, are calculated by equation (7). However, it has been proved through repeated experiments that the H.264 rate control algorithm may lead to a huge difference in peak signal-to-noise ratio (PSNR) at different initial PQs of the same video sequence, especially when the initial QPs are extremely small.

The PSNR, an objective evaluation indicator of video quality, can be calculated as:

$P S N R=10 \times \log \left(\frac{255^{2}}{M S E}\right)$   (8)

$M S E=\frac{1}{M \times N} \sum_{i=0}^{M-1} \sum_{j=0}^{N-1}(x(i, j)-y(i, j))^{2}$   (9)

where M is the width of the image; N is the height of the image; x(i,j) is the luminance or chrominance at (i,j) of the original image; y(i,j) is the luminance or chrominance at (i,j) after decoding. The unit of PSNR is dB.

$P S N R_{a v g}=\left(4 \times P S N R_{Y}+P S N R_{U}+P S N R_{V}\right) / 6$   (10)

where the variables PSNR_Y, PSNR_U and PSNR_V are the PSNR values of components Y, U and V obtained by the Joint Model (JM).

JM8.6 was adopted for our experimental platform. The parameter settings of the encoder in the JM test model are stored in the file encoder.cfg, and the corresponding parameter settings of the decoder are stored in the file decoder.cfg. In the rate control experiment, the encoder functions can be controlled by modifying the corresponding parameters in the file encoder.cfg. Some of the parameters in our experiment were configured as follows:

(1) FramesToBeEncoded=100;

(2) Framerate=30;

(3) Number Frames=0;

(4) Frame Skip=0;

(5) Search Range=16;

(6) InitialQP=15;

(7) RateControlEnable=1。

The first 100 frames in news_qcif and those in grandma_qcif were taken as the standard test sequences used in our experiment, with target bit rates of 128kbps and 64 kbps, respectively. The experimental results on the two sequences are listed in Table 1 below.

As shown in Table 1, when the QP was small (e.g. InitialQP=15), the actual code rate deviated from the target code rate by more than 21 %; moreover, the actual code rate and PSNR generated by the encoder fluctuated greatly with the InitialQP​s. This is attributable to the underestimation of the impact of GOP code rate on QP_st when the GOP layer rate control algorithm computed the initial QP QP_st . As a result, the initial QP of the I frame and the first P frame was not calculated accurately, leading to inaccurate bit number allocation to the GOP, which in turn affected the quality and code rate of the video.

The above problem was solved through the improvement below.

Step 1. Calculation of QP₀ 

When the encoder starts to process a video sequence, the first frame of the first GOP is treated as an I frame. Because it is the first frame, there is no reference information. Hence, the encoder uses the inherent QP. The initial QP QP₀  is still calculated by equation (5), where l₁, l₂ and l₃  are 0.1, 0.3 and 0.6, respectively.

Step 2. Calculation of QP_st 

According to the rate-distortion (R-D) model, the number of bits generated after the encoding of the i-1-th GOP, denoted as b_i-1, can be expressed as:

$b_{i-1}=M A D \times\left(\frac{a_{1}}{Q_{ {avestep}}}+\frac{a_{2}}{Q_{{avestep}}^{2}}\right)$   (11)

where Q_avestep is the quantization step length corresponding to the mean QP of the P frames in the i-1-th GOP.

Let Q_ststep be the quantization step length of QP_st . Then, the number of bits generated after the encoding of the i-th GOP, denoted as b_i, can be expressed as:

$b_{i}=M A D \times\left(\frac{a_{1}}{Q_{s t s t e p}}+\frac{a_{2}}{Q_{s t s t e p}^{2}}\right)$    (12)

$\alpha=\frac{b_{i-1}}{b_{i}}, \beta=\frac{a_{2}}{{ Q_{avestep} } \times a_{1}}$ and $\gamma=\frac{Q_{ {ststep}}}{ {Q_{avestep}}}$. Dividing equation (11) by equation (12), we have:  

$\gamma^{2}-\frac{\alpha}{1+\beta} \times \gamma-\frac{\alpha \times \beta}{1+\beta}=0$   (13)

Solving equation (13), we have:

$Q_{ {ststep}}=\frac{2}{\sqrt{1+\frac{4 \times(1+\beta)}{\alpha}}} \times Q_{ {avestep}}$   (14)

The QP QP_st  can be obtained according to equation (14).

Let QP_ave  be the QP corresponding to Q_avestep. Then, QP_st  must be subjected to the following constraint:

$Q P_{s t}=\min \left(\max \left(Q P_{a v e}-2, Q P_{s t}\right), Q P_{a v e}+2\right)$   (15)

where α is calculated during the determination of the number of bits needed for each GOP; β is initialized as 1 and updated constantly during the encoding process.

3.3 Experiment and analysis

The improved algorithm was applied in another experiment on JM8.6 under the same experimental conditions as the previous experiment. The experiment targets the standard test sequences of news_qcif and grandma_qcif, with target bit rates of 128kbps and 64 kbps, respectively. The experimental results of the improved algorithm are recorded in Table 2 below.

Table 2. Experimental results of the improved algorithm

Through in-depth analysis, the rate control technology in the H.264 standard was found to be unsuitable for the code rate features required for WSN channel transmission. Therefore, the original rate control algorithm was improved to enhance the stability and controllability of the code rate. The improvement mainly includes modifying the calculation method for QP_st  and enhancing the effect of the previous GOP on the QP_st  of the current GOP. Considering the cause of the high bit error rate of WSN channel transmission, the inaccuracy of the SMA was corrected by the BMA, and the secondary error concealment was introduced to overcome the shortcoming of the BMA, aiming to better utilize the completely concealed macroblock information and improve the quality of the reconstructed image. The improved algorithm was proved effective through experiments.