Flashcards
IVC
Flashcards do conjunto (75)
- 1HVS more sensitive to high contrast image regions than low contrast regions.Edges should be preserved. Artificial edge
artefacts introduced by quantisation are
very noticeable and should be avoided.
Spatial (contrast) masking of quantisation
noise can be exploited in textured regions. - 2HVS is more sensitive to luminance than chrominance information.Visual signal representations can be based
on colour spaces that split luminance and
chrominance. Chrominance channels can
be allocated lower bandwidth than the
luminance channel. - 3HVS is more sensitive to lower spatial frequencies than higher spatial frequencies.Frequencies beyond the upper range need
not be processed. Higher frequencies can
be coded more coarsely with implications
for quantisation. The lower frequency
luminance roll-off is not normally
exploited. - 4In order to achieve a smooth appearance of motion, the HVS must be presented with image frames above a certain minimum rate (and this rate depends on ambient light levels)Frame rates for acquisition and display
should normally be at least 50 Hz. Due to
flicker effects, higher frame refresh rates
are required for larger screens and closer
viewing. - 5HVS responses vary from individual to individual.Subjective quality assessment experiments
should be based on results from a large
number of subjects (typically > 20). - 6In video coding schemes it is usual to code the colour components in the form Y, Cb, Cr rather than R, G, B. Explain why this approach is justifiedThe human visual system has a relatively poor response to colour information.
The effective resolution of our colour vision is about half that of luminance
information. - 7I-frameIntra-frame.
Coded without any reference to any other picture. Used as a reference.
Does not exploit temporal correlation.
Errors do not progagate - 8P-frameInter-coded predicted frame.
Compressed difference between current frame and prediction of it.
Exploits temporal correlation but errors can propagate - 9B-frameBi-predicted pictures
Compressed difference between current frame and predidction of it based on up to two I, P or B frames - 10GOP in H.26412 Frames
1 I
3 P
8 B - 11Interlaced ScanningFields sampled at different times such that consecutive lines belong to alternate fields.
Doubles the percievd frame rate for the same bandwidth.
Artefacts with fast motion - 12Gamma CorrectionHVS not linear, content captured using devices is.
V_out = V_in^gamma - 133:2 Pulldown MotivationContent captured at 24fps.
Needs to be converted to 29.97 in the US, and 25fps in europe - 143:2 Pulldown Method1. Play film at 23.976 fps
2. Split frame into even/odd fields
3. At 23.976 there are 4 frames of film for every 5 frames of video - 15Bayer FilteringB G B G B
G R G R G
B G B G B
G R G R G - 16Normalised Bit Ratebits in compressed/bits in orgininal
- 17When can PSNR be deceptive?1. Phase shift in the reconstructed signal
2. Visual masking
3. Errors persist over time
4. Comparisions between different coding strategies - 18DSCQS Scoring5 Excellent
4 Good
3 Fair
2 Poor
1 Bad - 19Subjective Testing MethodologySequences in pairs
Ref/orginal in random order
MOS = arithmetic mean of scores
15 non-expert viewers
Screening
Diagram - 20Things that influence perceptual qualityInterest level
Display type
Lab conditions - 21Bit rate depends onEncoding algorithm
Content
Encoding parameters - 22RDOMaximise image/video quality dependent on constraints
- 23Three mechanisms of transformations1. Data decorrelation
2. Retained coeffcients can be quantised
3. Sparse matric remaining exhibits redundancy - 24When constructing a transform matrix, what must you remember?Make it orthonormal
- 25KLTProvides best energy compaction and yields optimal performance.
Transform is signal dependent therefore non-stationary - 26DCTGood approxiamtion to the KLT.
Derivation.
Complexity
Seperability - 27Why not use the DFT?1. Creates complex coeffcients, doubling the required data.
2. For short sequences where discontinuites are common, it can lead to ringing or spectral leakage in the frequency domain - 28DCT complexity (8x8) in MACSEach pass requires 8x64 MACs
2x8x64 = 1024 MACs - 29DCT complexityorder O(N^4)
- 30DCT complexity exploiting separabilityorder O(N^3)
- 31DCT complexity fft based methodO(N log2(N))
- 32McGoverns Algorithm - Reduction for 8x8Reduces multiplications required from 64 to 22
- 33Applications for lossless compressionMedical applications
Text or numerical information on a file - 34What do lossless methods employ to reduce bit rateSymbol statistics
- 35Self information-log2(P(A))
- 36When doing entrpoy encoding, and not given the probabilities, what can you do?Work them out from frequency of data
- 37Two limits of huffmanEcodes one symbol at a time.
Relies on pre-computed encode/decode stucture - 38Advantages of arithmetic encodingEach symbol may have a separate probability which may be pre-determined or calculated from previous data.
Symbol statistics can be adapted to the changing conditions
Performs better in the presence of channel errors - 39AC vs HC performanceAC 1dB better than HC at lower compression ratios
- 40Webers LawThe HVS is more sensitive to small changes in darker areas and much less sensitive in light areas.
- 41Motion can be translational, ...rotation, zoom and shear
- 42Disadvantages of higher order motion models1. Added complexity in the search.
2. Increased overheads and size of MV has to be increased - 43A translational motion model is used in practice because ...1. Low coding overhead
2. Simple to parameterise
3. At high frame rates, most motion is approximately rotational - 44Two methods used to deal with occulded objects1. B-frames, which use bi-directional prediction
2. MRF motion estimation - 45EBME vs BMEBME: SAD of inside of ref to outside of missing
EBME: SAD of outside for both ref and missing - 46Three types of motionGlobal
Local
Illumination - 47Choice of M,N for motion estimationSmall values give better predictions
But reduce the reliability since fewer pixels participate in the matching process. - 48SAD and SSD operationsSAD: 3MN
SSD: NM multiplications, NM-1 additions - 49BMME - effect of block sizeSmaller block sizes can deal better with complex motion result in better prediction at the cost of increased bit rates.
Larger is better for plain areas. - 50BMME - effect of search rangeA large search region can cope well with fast motions, at the cost of increased bit rate and a higher search complexity
- 51Error surfaceError surface can be multimodal containing one or more local minima
- 52Two methods to reduce complesity of BMME search1. Break loop if SAD is greater than previously found minimum,
2. Have a minimum acceptable BDM - 53Characteristics of a DFD compared to a conventional image1. Blocking artefacts caused by motion failure.
2. Artefacts around moving sharp edges resulting from small errors in MVs
3. Low correlation between adjacent pixels
THEREFORE: Decorrelating transforms will not perform as well on a DFD signal as the orginal intra-frame. - 54Why is a 4x4 integer transform used?Block size is matched to the basic motion estimation block, so it performs better in the presence of complex local motion.
Eliminates encoder/decoder mismatches.
Computationally efficient, can be implemented with only addition and shift operations. - 55Method to improve efficiency of MRF-MESimplex-based approach
- 56Where is it based to apply a deblocking filterWhen integrated into motion prediction loop to eliminate drift
- 57What is a waveletA mathematical function that provides an alternative means of splitting a signal into different frequency bands (subbands)
- 58Multirate theoryEnables a signal to be represented in the form of multiple band-pass components whilst maintaining a constant overall bit rate.
- 59Critical samplingSubsample the filter outputs in order to maintain a constant bit rate while preserving overall signal content
- 60Two network typesFixed IP network
The wireless network - 61Transmission typesDownloads and streaming
Interactive communication
Unicast
Multicast - 62ARQPackets checked via CRC
If error is detected a NACK is sent, and the packet is retransmitted - 63What happens if error rate exceeds the worst case scenario for FECCATASTRPHOE
- 64Intra-refreshI-frame inserted periodically to stop propagation of errors
- 65H.263 provides a 3-4dB improvement over H.261. How? (4)Half pixel motion compensation
Unrestricted motion vector mode
Advanced prediction mode
Syntax-based arithmetic coding - 66What is MPEG-4 part 2? Why did it not take off?Introduced new approaches using object based as well as waveform based methods.
To complex and R-D improvements were limited. - 67Additions to encoder architecture for H.264/AVCCoder control
Loop de-bocking filter
Intra-frame prediction - 68Skip ModeBlock SKIP means that no coding information is sent to the decoder. The motion vector for the skipped macroblock is based on the predicted MV only. No coded residual information is transmitted so only the motion compensated prediction is used to reconstruct the signal.
- 69Merge ModeBlock MERGE is similar to SKIP except that in this case adjacent blocks are 'merged'
so that they share the same motion vector. - 70For a filter bank F_0 = z^-1 x [n]={11011000...} y[n] = ?F_0 means a delay of one period.
Output is simply y[n] = {011011000...} - 71Header errors vs code word failureHeader failures can cause catastrophic error propagation.
Code-word bit errors can cause loss of code word and block synchronisation - 72Five methods to create an error resilient image1. ARQ
2. FEC
3. RVLC
4. FMO
5. EREC - 73Disadvantage of DCQS methodResults depend on person/lab conditions
Time consuming and expensive process - 74When does Huffman coding work best?When symbol probabilities are small
- 75On what type of video data will HEVC significantly outperform AVC and why?HEVC will outperform AVC particularly well on high spatial resolution formats due to its large CTU size and its flexible block partitioning.
Informações
- Descrição
- IVC FlashCards sobre IVC, criado por Josh Sowerby em 16-05-2018.
- Criado por
- Josh Sowerby
- Visualizações
- 13
- Avaliação