Method and apparatus for reducing mosquito noise in decoded video sequence

[0001] The present invention relates generally to image or video post processing, and more particularly to mosquito noise reduction in decoded video sequence.

[0002] Many popular image and video compression standards, such as JPEG and MPEG2, are based on Discrete Cosine Transform (DCT) processing. The basic approach of the DCT-based image and video compression technique is to subdivide the image into 8×8 blocks and then individually transform, quantize, and encode each block. In DCT-based compression techniques artifacts may occur, specially at low and moderate bit rates. Among several well-known artifacts, one is typically known as mosquito noise or ringing noise, which mostly appears in image homogeneous regions near strong edges. Mosquito noise is caused by loss of high frequency transform coefficients during the quantization step.

[0003] In recent years, several techniques have been proposed for mosquito noise reduction. Generally speaking, there are two types of techniques for mosquito noise reduction. One uses coding parameters to control the filtering process. The other does not rely on the coding information, and is usually called post-processing. The first type techniques require the decoder to send out some coding parameters along with the decoded video sequence, which in many practical situations are impossible. As such, applications for the first type of noise reduction techniques are limited. The second type techniques usually apply some adaptive filtering methods in order to avoid damaging image structures. Among the second type techniques, a technique proposed in the patent application WO 02/102086 uses a segmentation map to guide the filtering process. It classifies the image into five regions and assigns different regions with different filtering weights. FIG. 1 shows a block diagram of the image classifier 100 used in this method. However, classifying the image into five regions and using several cascaded filters, increases hardware complexity and cost.

[0004] The present invention addresses the above shortcomings. An object of the present invention is to provide a simple ringing (mosquito noise) region detection method that creates an accurate filtering map to guide the smoothing filter to effectively reduce mosquito noise.

[0005] Another object of the present invention is to provide a noise reduction apparatus for effectively reducing mosquito noise in decoded video sequence associated with the ringing region detection.

[0006] In one embodiment the present invention provides a ringing area detection method that classifies the input image into two regions: a mosquito noise region (i.e. filtering region) and a non-mosquito noise region (i.e. non-filtering region), and uses this classification information to adaptively remove the mosquito noise. Compared to aforementioned prior art, the present invention provides simpler ways of image segmentation for the purpose of mosquito noise reduction. The present invention further provides a method of avoiding abrupt changes between different regions and possible flicker, by providing soft-switching between filtering region and non-filtering region.

[0007] The present invention further provides a mosquito noise reduction system comprising a ringing area detector, a local noise power estimator, a smoothing filter, and a mixer. The ringing area detector includes an edge detector, a near edge detector, a texture detector, and a filtering region decision block. The ringing detection block detects the ringing area where the smoothing filter is to be applied. The local noise power estimator controls the filter strength of the smoothing filter. The smoothing filter smoothes the input image. The mixer mixes the smoothed image and the original image properly based on the region information from the ringing area detection block.

[0008] Other features and advantages of the present invention will be apparent from the following specifications taken in conjunction with the following drawings.

[0016] In the drawings, like reference numbers refer to similar elements.

[0017] FIG. 2 shows an example block diagram of a mosquito noise reduction system 200 according to an embodiment of the present invention. The mosquito noise reduction system 200 comprises a ringing area detection unit 202 including a gradient computation unit 204, an edge detection unit 206, a near edge decision unit 208, a texture detection unit 210, and a filtering region decision unit 212. The mosquito noise reduction system 200 further comprises a local noise power estimator 216, a smoothing filter 218, and a mixer 220. The local noise power estimator 216 includes a high pass filter (HPF) 222, a local standard deviation calculation unit 224 and a converter 226. The function of the ringing area detection unit 202 is to detect the ringing region (or mosquito noise region) where the smoothing filter 218 needs to be applied.

[0018] The mosquito noise reduction system 200 can be applied to both the luminance signal Y and the chrominance signals U and V in a video signal. As such, in general, the input signal is denoted by F(i,j), where i and j are the indices for image row and column respectively. Referring to FIG. 2, the gradient computation unit 204 computes the gradient of the input signal F(i,j), generating the gradient |∇F(i,j)|. Those skilled in the art will recognize that different numerical methods can be used to calculate the gradient ∇F(i,j) for a given image.

[0019] Using the gradient ∇F(i,j), edges and texture information are then examined. The edge detection unit 206 detects an edge by comparing the magnitude of the gradient, |∇F(i,j)|, with a threshold value T₁ as:

[0020] $E\left(i,j\right)=\{\begin{array}{cc}1,& \mathrm{if}\nabla F\left(i,j\right)>T_1\\ 0,& \mathrm{else}\end{array}$

[0021] where 1 denotes edge, and 0 denotes non-edge, and T₁ is a predetermined constant.

[0022] Because the ringing artifacts in compressed video arises near edges, it is necessary to detect whether a current pixel is near an edges or not. For this purpose, the near edge detection unit 208 checks whether there are edges in the neighborhood of the current pixel. Specifically, considering a P×Q image pixel block and counting the edge samples in this block, if the block contains more than a certain number of edges, it is assumed that the current pixel is located near edges. Accordingly, let N_e^P×Q(i,j) denote the number of edge samples in the P×Q block around the current pixel (i,j), then the near edge detection NE(i,j) is computed as:

[0023] $\mathrm{NE}\left(i,j\right)=\{\begin{array}{cc}1,& \mathrm{if}{N}_e^{P\times Q}\left(i,j\right)>N_T\\ 0,& \mathrm{else}\end{array}$

[0024] where 1 denotes near-edge pixel and 0 denotes non-near-edge pixel, where N_T is a predetermined constant.

[0025] For simplification, the present invention also proposes to use non-overlapped blocks for near edge detection. (Note that the block for computing N_e^P×Q(i,j) assume overlapped blocks or sliding blocks as pixels are processed.) That is, as shown by example in FIG. 3, the image 300 is divided into multiple P×Q blocks, and then the near-edge block NEB (I, J) is detected as:

[0026] $\mathrm{NEB}\left(I,J\right)=\{\begin{array}{cc}1,& \mathrm{if}{N}_e^{P\times Q}\left(I,J\right)>N_T\\ 0,& \mathrm{else}\end{array}$

[0027] where N_e^P×Q(I,J) represents the number of edge samples of the (I,J)th block in which I=└i/P┘, J=└j/Q┘, and └x┘ is the floor function which returns to the integer part of the nonnegative number x . Further, NEB(I,J)=1 implies that the (I,J)th block is a near-edge block. Then, the near edge detection for the current pixel F(i,j) is set as:
NE(i,j)=NEB(I,J).

[0028] Note that with the use of non-overlapped block, the pixels that belong to the (I,J)th block have the same near-edge information. This significantly reduces computation complexity when computing near-edge detection compared to the overlapped block based detection where it is assumed that near edge detection needs to be performed for every pixel position repeatedly.

[0029] Referring back to FIG. 2, the texture detection unit 212 detects the texture TX(i, j) by comparing |∇F(i,j)| with another threshold value T₂, as:

[0030] $\mathrm{TX}\left(i,j\right)=\{\begin{array}{cc}1,& \mathrm{if}\nabla F\left(i,j\right)>T_2\\ 0,& \mathrm{else}\end{array}$

[0031] where 1 denotes texture, 0 denotes non-texture and T₂ is a predetermined constant which is less than T₁.

[0032] The filtering region decision unit 214 generates the filtering region map R(i,j) based on the supplied near-edge map NE(i,j) and the texture map TX(i,j). The near-edge map actually already marks the image region where the mosquito noise occurs. However, the near-edge map may also contain some texture and filtering texture will cause blurring. Therefore, in order to obtain an accurate filtering region map, texture needs to be eliminated from the edge-block map. The logic for removing the texture from the near-edge map is:

[0033] $R\left(i,j\right)=\{\begin{array}{cc}1,& \mathrm{if}\mathrm{NE}\left(i,j\right)=1\mathrm{and}\mathrm{TX}\left(i,j\right)=0,\\ 0,& \mathrm{otherwise};\end{array}$

[0034] where 1 denotes the filtering region where ringing artifact occurs and 0 denotes the non-filtering region.

[0035] Since the near edge decision unit needs P lines of image data to generate the near-edge map, line memory delay may happen at the filtering region decision unit 214. The filtering region map generated by the filtering region decision unit 214 is then supplied to the mixer 220.

[0036] The purpose of the local noise power estimator 216 in FIG. 2 is to provide the local filter strength σ_n to be used by the smoothing filter 218. The HPF 222 extracts mainly the noise component of the input signal. In a preferred embodiment, the following HPF is used:

[0037] $\left[\begin{array}{ccc}1& -2& 1\\ -2& 4& -2\\ 1& -2& 1\end{array}\right].$

[0038] The local standard calculator 224 calculates the local standard deviation σ_h of the high-pass filtered signal HF(i,j) over a r×s window in the video image as:

[0039] ${\sigma }_h\left(i,j\right)=\frac{1}{r\times s}\sum _{n=-r}^r\sum _{m=-s}^sHF\left(i+m,j+n\right)-{\mu }_h\left(i,j\right),\mathrm{where}$ ${\mu }_h\left(i,j\right)=\frac{1}{r\times s}\sum _{n=-r}^r\sum _{m=-s}^{s}HF\left(i+m,j+n\right).$

[0040] The calculated local standard deviation σ_h is forwarded to the converter 226 which converts the local ringing noise power σ_h to the equivalent local additive Gaussian noise power σ_n. The converter 226 implements the following relation:
σ_n=Bσ_h^0.45,

[0041] where B is a user controlled global filter strength parameter which controls the overall smoothing level of the smoothing filter. The estimated local noise power σ_n is supplied to the smoothing filter 218.

[0042] The smoothing filter 218 comprises an edge preserving filter which removes noise while retaining image edges. The output F_NR(i,j) of the smoothing 218 filter is supplied to the mixer 220. There are many examples of edge preserving filters that can be used as the smoothing filter 218 in the mosquito noise reduction system 200 according to the present invention. One example smoothing filter is the weighted sigma filter, which is defined as:

[0043] $F_{\mathrm{NR}}\left(i,j\right)=\frac{1}{N}\sum _{n=-r}^r\sum _{m=-r}^r{w}_{m,n}·F\left(i+m,j+n\right);$ $\mathrm{wherein},\text{}{w}_{m,n}=\{\begin{array}{cc}2& \mathrm{if}F\left(i+m,j+n\right)-F\left(i,j\right)<C_1{\sigma }_n,\\ 1& \mathrm{if}F\left(i+m,j+n\right)-F\left(i,j\right)<C_2{\sigma }_n,\\ 0& \mathrm{otherwise};\end{array}N=\sum _{n=-r}^r\sum _{m=-r}^r{w}_{m,n};$

[0044] and C₁, C₂ are predetermined constants with 0<C₁<C₂, and σ_n is the local noise power.

[0045] Another example smoothing filter is the minimal mean square error (MMSE) filter. The MMSE filter is defined as:

[0046] $F_{\mathrm{NR}}\left(i,j\right)=\mu \left(i,j\right)+\frac{\max \left({\sigma }^2\left(i,j\right)-{\sigma }_n^2,0\right)}{\max \left({\sigma }^2\left(i,j\right)-{\sigma }_n^2,0\right)+{\sigma }_n^2}·\left[F\left(i,j\right)=\mu \left(i,j\right)\right],$

[0047] wherein σ_n is the local noise power, and, μ(i,j) and σ(i,j) are the local mean and local standard deviation computed over a r×r window as:

[0048] $\mu \left(i,j\right)=\frac{1}{r^2}\sum _{n=-r}^r\sum _{m=-s}^{s}F\left(i+m,j+n\right),\sigma \left(i,j\right)=\frac{1}{r^2}\sum _{n=-r}^r\sum _{m=-s}^sF\left(i+m,j+n\right)-\mu \left(i,j\right).$

[0049] The mixer 220 in FIG. 2 selects either the output F_NR(i,j) of the smoothing filter 218 or the original input signal F(i,j) as the final output based on the filtering region map, separating the image into two regions: filtering region and non-filtering region. In the filtering region, the output F_NR(i,j) of the smoothing filter 218 is selected as the final output. In the non-filtering region, the original input signal F(i,j) is selected as the final output.

[0050] There are many example variations of the mosquito noise reduction system 200 shown in FIG. 2, according to embodiments of the present invention. FIG. 4 shows a block diagram of a system 400 which is a variation of the mosquito noise reduction system 200 shown in FIG. 2, wherein the system 400 additionally uses the edge map for filtering region decision. In this case, edges are included in the filtering region map. With the supplied edge map E(i,j), the filtering region decision unit 214 in the system 400 generates the filtering region map as:

[0051] $R\left(i,j\right)=\{\begin{array}{cc}1& \mathrm{if}\left(\mathrm{NE}\left(i,j\right)=1\mathrm{and}\mathrm{TX}\left(i,j\right)=0\right)\mathrm{or}\left(E\left(i,j\right)=1\right),\\ 0& \mathrm{otherwise};\end{array}$

[0052] where 1 denotes the filtering region and 0 denotes the non-filtering region.

[0053] FIG. 5 shows a block diagram of an example system 500 which is yet another variation of the mosquito noise reduction system 200 shown in FIG. 2, that additionally uses binary filters (BF) 201 to filter the edge map and the texture map. The purpose of using the binary filters 201 in the system 500 is to remove the small holes in the edge map and the texture map. Preferably, the type of binary filters 201 used is the mathematical morphology closing operator. If g(i,j) is a binary function, then the closing operator C_w(·) is defined as:
C_w(g)(i,j)=min_w(max_w(g(i,j))),

[0054] wherein W is a window centered at the pixel (i,j), and min_w(·) and max_w(·) are the functions that return the minimal and maximal values within the window W, respectively. Preferably, the window W used for the closing operator is a 5×5 window.

[0055] FIG. 6 shows a block diagram of an example system 600 which is yet another variation of the mosquito noise reduction system 200 shown in FIG. 2, wherein the system 600 uses a soft switching unit 230 to control the smoothing filter 218 of the system 600. In the system 600, the filtering region map R(i,j) first passes through a low-pass filter 203 to generate a smoothed version of the filtering region map {tilde over (R)}(i,j). The function of the low-pass filter 203 is to create a smooth transition between the filtering region and the non-filtering region. With this smoothed region map, hard switching can be replaced by soft switching. Therefore, abrupt changes between the filtered region and the non-filtered region in the processed image can be avoided using the system 600.

[0056] In FIG. 6, the soft switching unit 230 uses the smoothed filtering region map {tilde over (R)}(i,j) to control the switching between the result of the smoothing filter F_NR(i,j), and the original signal F(i,j). The basic idea is to use the filter result F_NR(i,j) in the filtering region (i.e., mosquito noise region) and gradually switching to the original signal when getting into the non-filtering region (i.e., non mosquito noise region). The soft switching unit 230 determines the final output of the systemF_OUT(i,j) as:
F_OUT(i,j)={tilde over (R)}(i,j)·F_NR(i,j)+(1−{tilde over (R)}(i,j))·F(i,j).

[0057] Since the near edge decision unit map in a mosquito noise reduction system according to the present invention needs P lines of image data to generate the near-edge map, for line memory constrained hardware implementation, system delay may happen. Accordingly, referring to FIG. 7, a mosquito noise reduction system 700 using delayed near-edge map is provided to reduce the hardware complexity. The system 700 is yet another variation of the mosquito noise reduction system 200 shown in FIG. 2, wherein system 700 in FIG. 7 uses a delayed near-edge map stored in a memory unit (M) 240 to reduce the hardware complexity. In this case, the filtering region decision unit uses the delayed near-edge map, which is generated from previous frame and is stored in the memory unit 240, and the texture map of the current frame to generate the filtering region map. Further, the near-edge map of the current frame is passed to the memory unit 240 instead of the filtering region decision unit 214. The other components of the system 700 are similar to those of the system 200 shown in FIG. 2.

[0058] The present invention has been described in considerable detail with reference to certain preferred versions thereof; however, other versions are possible. Therefore, the spirit and scope of the appended claims should not be limited to the description of the preferred versions contained herein.