Questions for consultation: Please write your questions here.
Before asking organizational questions, read the Exam FAQ in Opal:
https://bildungsportal.sachsen.de/opal/auth/RepositoryEntry/3190226946/CourseNode/101900007036152?8
Organizational Questions:
- Q: In terms of notation of decimal point, should we use english/american (. - point) or german (, - comma) in the exam?
- A: If you answer in English, use point. If you answer in German, use comma.
- Q: Are we allowed to use any color of pen for the exam? Is red/green allowed?
- A: No. Please use black or blue.
- Q: Upto how many digits after decimal should the solution be solved? (for calculations)
- Q: How many points paper will be in total?
- Q: What will be the total time of the paper?
- A: 60 min officially. We will give you 10 minutes extra.
- Q: Müssen wir die Formeln für RGB -> YCbCr, HSV kennen?
- A: No, these equation would be given, if required. You should know equations for entropy, information content, file sizes, error metrics, ...
- Q: Also for RGB -> CMY and formula to calculate Paethpred? Do we need to memorize?
Subject-related Questions:
- Q1: Relating to the calculation of entropy for a message: How are the probabilities taken? For example, for Task 5 of Tutorial 1, the set of characters apparently is evenly distributed and we calculate then the probabilities according to the number of times that it appears in the word. What if we are not given a word, but a set of characters that is not evenly distributed? Should we then use the probabilities from the character set only?
- Q2: For the construction of the code tree using Shannon-Fano: Are the groups always made from most frequent to least frequent? Is it possible to make a group with the highest and smallest frequency characters in order to make the groups as balanced as possible?
- A: No. The first step in Shannon-Fano is always to sort the characters in a descending way.
- Q3: With regards to video encoding: How do P-frames actually work? If the motion vectors are looking for the smallest MAD, how does the image then not repeat itself? In this case MAD would be 0.
- A: In p-frames in fact the image really repeats parts (macroblocks) of the previous frame by just encoding a motion vector pointing to the most silimar block in the previous frame in older mpeg versions. in newer mpeg versions also differenes can be encoded. It can happen that MAD = 0 if the previous image contains completely identical image regions, e.g. in cartoons. In real movies, it is very unlikely, that the MAD gets 0, since we have camera noise between two subsequent frames, even if there is no apparent motion in the image.
- Q4: What points to mention in difference between JPG, PNG, BMP, GIF?
- A: Answer was explained live. Please refer back to the tutorial slides and lecture notes.
- Q5: What are disadvantages of vector graphics?
- A: They are not suited for real-world photographies. The file size can blow up, if the mathematical description of your image content is complex.
- Q6: In slide 13 tutorial 9, where is auditory masking threshold?
- A: Was explained live during the consultation session.
- Q7: What is an application of HSV? Can digital camera be an application, too?
- A1: Artistic manipulations of images, photo editing, color filters, object detection, etc.
- A2: Yes, digital camera would be a valid application, too.
- Q8: TUTORIAL 7 task2 mentions GIF without dithering..is that true?
- A: Dithering is an image processing concept. In that regard it is not really limited to a specific image format. You could also apply dithering to an image and save that image as JPEG. It's just that GIF already has the dithering procedure built-in as a preprocessing step.
- Q9: Do we need to memorize bitstream message format of GOB?
- Q10: Could you please summarise main characteristics of BitMap and its area of application? Can we say BMP supports transparency, if (as I saw) it is only supported in 32bit encoding option, but not with smaller?
- Q11: Could you please summarise steps of lossless JPEG compression?
- A: Color conversion RGB -> YCbCr, Chroma Subsampling with 4:4:4, Calculation of DCT, Quantization Tables with all '1', ZigZag Encoding and Huffmann Encoding
- Q12: I mean: In lecture there was a different one JPEG-Lossless, with prediction technique. See 3.3.1, Lecture 5.
- A: Please, see question Q23.
- Q13: How do we calculate log values without using calc?
- A: log_2 (2) = 1, log_2 (4) = 2, log_2 (8) = 3, ....
- Q14: How to interpret this formula: 𝑆𝑠𝑟𝑐 = (𝑁𝐴/𝑁𝐵) ≥ 𝐻𝑠𝑟𝑐 [𝐵𝑖𝑡𝑠 / 𝑊𝑒𝑟𝑡] on page 17 of the script? (page 16 in english script; page number refers to page number that is printed on the page, not the number the reader outputs)
- A: N_a is the total size of a signal in Bits (e.g. 2400 Bits), N_b is the number of values in the signal (e.g. 100 pixels). So S_src will give 24 Bits / value in this example.
- Q15: Is the example (in gray box) on page 30 (29 in english script) of the script for MSE calculation wrong? It does not divide by number of elements...
- A: It should be divided by the Number of N=3. The division by N is indeed missing there.
- Q16: How do we calculate log with base 2 without calculator?
- Q17: Can the codes of Huffman or Shannon-Fano be diffrent and its still correct if the code length are the same?
- A: Yes, if multiple interpretations are possible, we can allow multiple solutions (if they are consistent).
- Q18: Should we expect more problems or theory questions?
- Q19: Tutorial 2, task 8: What is the difference between quantizing a signal and reconstructing a signal? Is quantizing "x to q mapping"and reconstruction "q to x mapping"?
- A: Quantizing means converting a value-continouus signal into a value-discrete signal. This happens during digitization of an analog signal. Reconstruction converts the digitized signal back to an analog signal.
- Q20: Tutorial 2, task 8: For quantizing x=2, is the value for q in range [1,3]? Can the value of q be in decimals, e.g. 1.5?
- A: No. You have to work with half-open intervals here, i.e., a definition has to be made how to deal with the edges/boundaries. I would suggest intevals like. [-4, 0), [0, 2), [2, 4), etc. q has to be integer.
- Q21: Are there 4 motion detection techniques? Full search, area search, logarithmic, heirarchical?
- Q22: Can you please repeat what hierarchical is?
- A: Was explained live in the consultation session.
- Q23: It was mentioned, that jpeg also use lossless encoding, but we always mentioned in the comparission, that one difference between for example png is the lossly encoding. So is it just not common to use lossless encoding on jpeg or why we always mentioned it. Furthermore, I didn't really got the difference between jpeg lossless and png lossless. But in 3.3.1 in the script the lossless encoding of jpeg is described? Not in the way you currently describing
- It is right jpeg-ls will not be part of the exam. The method indeed works differently from what we described in Q30. It is a seperate defintion of jpeg without the 4 main steps, which just relies on the encoding described in section 3.3.1
- Q25: Bitmap-Info-Header --> Difference between Farbtiefe and Bit/Pixel? Page 63 in the german script, Abbildung 3.4c. Difference of biPlanes and biBitCnt. For me it expresses the same information.
- A: "biPlanes" is a field in the BMP file header that specifies the number of color planes used in the bitmap image. "biBitCount" is another field in the BMP file header that indicates the number of bits per pixel used in the bitmap image.
- Q26: Can we revisit tutorial 2 task 1 (how do we make sure f-sample>2fmax when drawing the reconstructed signal).
- A: When drawing the recosntructed signal, just connect the sample points with a smooth curve, without any turning points in between.
- Q27: In task 8-a&b.I didn't get how we are quantizing the non-uniform midrise quantizer ,I read that we have to use companding or Lloyd max algorithm.Is it possible to just map x to q?
- A: Just use the graph to quantize the signal.
- Q28: How to mark dead zone and amplitude clipping in question Task 8 Tutorial 2 from mid-tread function graph?
- Just mark it, e.g. by drawing a circle around it.
- Q29: Could you explain the compression types of Bitmap (RLE8, RLE4)? Page 63 in the german script, Bitmapinfoheader (biCompression).
- A: Sounds like Run length encoding which is a applied to images with 4 bit or 8 bit color palette.
- Q30: Are these answers correct? Explain the principle or configuration of JPEG compression, which allows lossless encoding. Chroma Subsampling with 4:4:4 encoding, Applying Discrete Cosine Transform to small (8x8) pixels blocks for the complete image, Quantization using lookup tables, which consist of only “ones” (in fact this means no quantization), Compression through a lossless encoding like Run Length- or Huffman encoding
- Describe the basic steps of the lossy JPEG compression. 1) Data reduction using chroma subsampling, 2) Discrete cosine transformation on small block size, 3) Quantization using lookup tables, 4) Compression by Hufmann algorithm
- Q31: Calculate the PaethPred value for the pixel X using the Paeth Predictor algorithm
- A: Please see solution slide 8 of Image Encoding III. Follow the flow chart.
- Q32: Summarise please main characteristics of BitMap and for which usecause this format is used?
- A: It is losslessly encoded, supports high color depth --> used if not many images are created (and thus memory does not play a big role), but quality should be preserved (medical images, sketches/diagrams with > 256 colors).
- Q33: Could you please explain tutorial 6 task 2.
- A: Was explained live during the consultation session.
- Q34: Wrt to the lossy JPEG compression, is the data reduced during quantization, am I correct?
- Yes. The values of the DCT coefficients are divided down, leading in some cases to values close to 0, which will then be rounded down to 0. The string of zeros can then be compressed well.
- Q35: While compressing using quantization table, do we have to strongly quantize the low frequencies (top-left) or high frequencies (bottom-right)?
- A: The higher frequencies live in the bottom right. Those are the ones which can be quantized a lot, without losing a lot of visible information.
- Q36: In huffman encoding, how we choose letter if its value is the same like c=4,h=4,E=4 and should we maintain it its right side tree then always left tree is smaller than right side tree.
- A: You are free to decide. Please, just do it in a consistent way, e.g., always sort combined nodes back into the table above a previous entry (with the same sum), or below it.