Upload
yara-ali
View
83
Download
2
Embed Size (px)
Citation preview
Visual Quality Enhancement in DCT-Domain Spatial Downscaling
Transcoding Using Generalized DCT Decimation.
Presented by:
Marwa AhmedMona Ragheb
Sara SeragYara Ali
Agenda
• Introduction Definition What is meant by:
Transcoding Spatial Domain & DCT Domain Downscaling Alias Quantization
Video Adaptation
Frequency Synthesis
• Generalized DCT Decimation For Spatial Downscaling
Agenda
• Computation Reduction Using Sparse matrix representation
• Analysis of the proposed DCT decimation filter
• Experimental results Optimal list squares Up-Scaling filters ( Steps )
• Peak Signal to Noise ratio
• Conclusion
Visual Quality Enhancement in DCT-Domain Spatial Downscaling Transcoding Using Generalized DCT
Decimation.
• The goal image enhancement is to improve the image quality so that the processed image PSNR is high and less computational complexity
Abstract
1. we propose a generalized discrete cosine transform (DCT) decimation scheme for DCT-domain spatial downscaling which performs two-fold decimation on subframes of a flexible size larger than the traditional
8 ᵡ8 block size to improve the visual quality.
2.Efficient sparse-matrix:representations are then derived to reduce the
computation of the proposed DCT decimation method.
Abstract (cont.)
3.We comparethe filtering performances and computational complexities of the
proposed scheme and the pixel-domain downscaling schemes Our analysis shows that :
proposed scheme can reduce the aliasing artifact compared to the pixel-domain downscaling schemes,
Where as the computational complexity may be increased We also: integrate the proposed decimation scheme into the cascaded
DCT-domain transcoder for spatial downscaling of a pre encoded video into its quarter size
Experiments show the proposed approach can achieve better visual quality than the existing schemes
What is meant by: Transcoding : Video transcoding is an
operation of converting a video bit-stream into from one format into another format (e.g., bit-rate , frame-rate, spatial resolution, and coding syntax). It is an efficient means of achieving fine
and dynamic video adaptation. Video adaptation : convert the video bit
rate according to the channel conditions. Since the pre-encoded video is encoded at high quality and bit rate. For low bandwidth connections, the video bit rate needs to be converted to low bit rate.
Spatial Domain & DCT Domain
Spatial Domain (Image Enhancement):
Definition:is manipulating or changing an image representing an object in space to enhance the image for a given application.• Techniques are based on direct manipulation of pixels in an
image• Used for filtering basics, smoothing filters, sharpening filters,
unsharp masking and laplacian
Discrete Cosine Transform (DCT) domain
• This allows us to discard those equations involving the higher frequency components, reducing the size of the equation set considerably.• in the DCT domain, each equation’s significance is
dependent on the corresponding DCT frequency• Does not affect the compressibility of the original image
because it enhance the image in the decompression
Aliasing:• When a signal is under-sampled, aliasing can result• Aliasing is when high frequency components masquerade as lowfrequency ones, and can result from improper image sampling
Downscailing : The operation of retaining the low-frequency coefficients of aDCT sub-frame and taking the half-size IDCT Each N×M sub-frame is extracting only the (N/2)×(M/2)low-frequency.
Quantization: is the process of converting a continuous analog audio signal to a digital signal with discrete numerical values.
Frequency synthesis : downscaling method first synthesizes an incoming macroblock consisting of four 8 ᵡ 8 DCT blocks intoone 16 ᵡ 16 DCT block, and then obtains the downscaled 8 ᵡ 8DCT block by extracting the 8 ᵡ 8 low-frequency DCT coefficients of the 16 ᵡ16 DCT block
• In realizing a transcoder, the computational cost and the picture quality are usually the two most important concerns.
A cascaded DCT-domain transcoder (CDDT): as depicted in Fig. 1, was first proposed in for spatial downscaling where a DCT-domain bilinear filter was used as the anti aliasing filter for the spatial downscaling.
• cascade a decoder followed by an encoder. This cascaded pixel-domain architecture is flexible and can be used for bit rate adaptation and spatio-temporal resolution conversion without drift.
MC: reduces the temporal redundancy.
DCT: reduces the spatial redundancy and achieves energy compaction
Quantization is performed to achieve higher compression ratio. Variable-length coding.
VLC: is applied after the quantization to reduce the remaining redundancy. decoder decodes the compressed input video encoder re encodes the decoded video into the target format
A video picture is predicted from its reference pictures and only the prediction errors are coded.
the encoder reuses the motion vectors along with other information extracted from the input video bit stream.
II. Generalized DCT decimation for spatial
downscaling
Formulation of generalized DCT decimation
Formulation of generalized DCT decimation
STEP 1:
A group of consecutive 8-samples DCT vectors are first transformed into an N-pixel vector by 8-point IDCT, Where N is a multiple of 8.
Formulation of generalized DCT decimation
The N-pixel vector is then transformed into its corresponding DCT vector by N-point DCT
Formulation of generalized DCT decimation
The N-point DCT representation of fN can be computed by:
fN: N-pixel vector that’s composed of 8-pixel vectors bi , i= 1……, N/8
TN: N-point DCT transform matrix that’s divided into N/8 columns of submatrices TN,i of size Nx8
Formulation of generalized DCT decimation
Formulation of generalized DCT decimation
STEP 2:
DCT decimation is subsequently performed on the N-sample DCT vector by extracting the N/2 low-frequency DCT coefficients followed by N/2-point IDCT to obtain a downscaled N/2-pixel vector
Formulation of generalized DCT decimation
Formulation of generalized DCT decimation
STEP 3:
The N/2-Pixel vector is transformed into a group of consecutive 8-Sample DCT vectors by 8-Point DCT to form the output DCT array
Formulation of generalized DCT decimation
•Computation reduction using sparse matrix representations
•Analyses of DCT-DECIMATION downscaling filters
COMPUTATION REDUCTION USING SPARSE MATRIX REPRESENTATIONS
To reduce computation for matrix operations in (4) and (7)
can be represented in sparse matrix form
COMPUTATION REDUCTION USING SPARSE MATRIX REPRESENTATIONS The following characteristics have been noted in with dimension (N/2) * 8:
1. General case:The entries of r th row in are all zeros except the r th entry
where r = 0, N/8, 2N/8, 3N/8About N/8 of the entries are zeros.2. Special case 1:For K = N/8 is even
Where i = 1, …….., N/16 r = 0, ……, N/2 and c = 0, …, 7 3. Special case 2:for K = N/8 is odd
Where i = 1, …….., N/16 r = 0, ……,, N/2 and c = 0, …, 7for matrix with i = N/16 + 1 , the entries of odd values of r + c is zero for
r ≠ 0, N/8, 2N/8 , 3N/8at most half of the entries are zeros.
COMPUTATION REDUCTION USING SPARSE MATRIX REPRESENTATIONS
Based on the previous facts: are defined to reduce computations ,For i = 1, …., k/2 where k is even :
Substituting in :
ANALYSES OF DCT-DECIMATIONDOWNSCALING FILTERS
The operation of retaining the low-frequency coefficients of a DCT sub-frame and taking the half-size IDCT is, in effect, to perform anti-aliasing filtering and then followed by downsampling on the sub-frame in the pixel domain.
Following is the analysis of the performances and complexities of various downscaling filters for the 1-D case.
For N samples of 1-D signal x, when downscaled by a factor of two, the downscaled N/2-sample signal y is obtained as follows:
The downscaling filter is defined as :
Which is considered as a linear filter.
ANALYSES OF DCT-DECIMATIONDOWNSCALING FILTERS
The linear transform can be represented as an N-band filter bank structure
the z-transform of the output y can be obtained by:
As N increases, the gain of DCT decimation filters, |F0(z)|, becomes much flatter in the low frequency part (0~π/2), whereas the gain decreases rapidly in the high-frequency part (π/2~π).For the bilinear filter, the gain in the high-frequency part is always larger than its counterparts of DCT decimation filter and 7-tap filter.The smaller gain in the high-frequency part implies less visible aliasing artifacts in the downscaled image.
the magnitude responses of the two pixel-domain filters: the bilinear filter and the 7-tap filter, and the generalized DCT decimation filters with N = 8, 16, 72, and 288
• increasing the sub-frame size for the DCT decimation filters will lead to better quality of downscaled image but it will also increase the computational complexity significantly.
• The shown table lists computational complexities with different N values and different filter:
N Avg. Multiplication
s
Avg. additions
Gen. DCT
decimation
Sparse
matrix deco
mposition
Gen. DCT
decimation
Sparse
matrix deco
mposition
352 4224 2788 4208 2792
32 384 228 368 232
16 192 100 176 104
8 64 20 64 20
Tab length Avg. Multiplicat
ions
Avg. additions
Bilinear 8 8
7-tab Gaussian
56 48
Average computational complexity using generalized DCT-decimation scheme
Average computational complexity for pixel-domain downscaling scheme
Experimental Results
In our experiment:
We use: One CIF (352* 288) Two ITUR(704*576) In each video 150 frame Encoded by front-end MPEG-2
encoder Each coded video is transcoded by
using CDDT
Resulting in a spatially downscaled video of quarter size
We implement bilinear filter and 7-tap gaussian filter
Each downscaled image is decoded and up-scaled to its original size
The optimal least- squares upscaling filter matrix minimize the error between original sized image & its reconstructed (downscaled & then upscaling)
Steps: Divide each downscaled pixel vector into N/2 sample pixel vector then
transform it to N/2 sample DCT vector.
Expand the size of each N/2 sample DCT to N sample by padding zero coefficient in high frequency bands
Apply N-point IDCT
We compare PSNR values of o/p image of downscaling filter followed by upscaling filter & final o/p of transcoders.
PSNR, is an engineering term for the ratio between the maximum possible power of a signal and the power of corrupting noise .
So if PSNR is high the noise is low so the quality is high
PSNR= 20 log10 (255/ RMSE)
Conclusion: We experiment results in N=8,16,32
In N=16,32 we have better visual quality over that N=8 but in the same time we increased computational complexity
So we can use N=8 in low activity region & N=16,32 in high activity region to achieve good trade-off between computational complexity & visual quality.
Thank You !