Transcript Slides
CMSC 100 Storing Data: Huffman Codes and Image Representation Professor Marie desJardins Tuesday, September 18, 2012 CMSC 100 -- Data Compression 1 Tue 9/18/12 Data Compression: Motivation Memory is a finite resource: the more data we have, the more space it takes to store Same with bandwidth: the more data we need to send, the more time it takes Data compression can reduce space and bandwidth Lossless compression: Store the exact same data in less space Lossy compression: Store an approximation of the data in less space 2 CMSC 100 -- Data Compression Tue 9/18/12 Time and Space Tradeoffs Data compression trades (computational) time for space and bandwidth: It takes time to convert the original data D to the compressed format DC It takes time to convert compressed data DC back to a viewable format D’ Compression ratio: Length(DC ) CR Length(D) 3 Space savings: CMSC 100 -- Data Compression SS 1 CR Tue 9/18/12 Lossless vs. Lossy Compression Lossless: Save space without losing any information Take advantage of repetition and self-similarity (e.g., solid-color regions in an image) Lossy: Save space but lose some information Lose resolution or detail (e.g., “pixillate” an image or remove very high/low frequencies in a sound file) 4 CMSC 100 -- Data Compression Tue 9/18/12 Encoding Strategies Run-length encoding: replace n instances of object x with the pair of numbers (n,x) Frequency-dependent encoding: use shorter representations (fewer bits) for objects that appear more frequently in a document Relative or differential encoding: when x is followed by y, represent y by the difference y-x (which is often small in images etc. and can therefore be represented by a short code) Dictionary encoding: Create an index of all of the objects (e.g., words) in a document, then replace each object with its index location (can save space if there is a lot of repetition) 5 CMSC 100 -- Data Compression Tue 9/18/12 Image and Sound Formats Images Sound 6 Row-by-row bitmaps in different color spaces: RGB (one byte per color = 24 bits = 17M different colors), a.k.a “True Color” (used in JPEG formats) (How much storage for one True Color 2Kx3K digital camera image?) Color palette: Use only one byte to index 256 of the 17M 24-bit colors (used in GIF formats) (How much storage for one 24-bit color 200x300 image on a website?) Variable resolution provides different image sizes and levels of fidelity to an original (continuous or very high-resolution digital) image Convert continuous sound to digital by sampling (variable-rate) Each sample can be represented with varying levels of resolution (“bit depth”) (MP3: 44K samples/second, 16 bits/sample – how much storage for one minute of sound?) CMSC 100 -- Data Compression Tue 9/18/12 Compression Ratio: Example Suppose I have a 2M .PNG (bitmap) image and I store it in a compressed .JPG file that is 187K. What is the compression ratio? What is the space savings? 7 CMSC 100 -- Data Compression Tue 9/18/12 Huffman Coding Lossless frequency-based encoding Huffman coding is (space-)optimal in the sense that if we need the exact distribution (frequency) of every object, we will be able to represent the document in the shortest possible number of bits Downside: It takes a while to compute Goal #1: Length of each object should be related to its frequency 8 Specifically: length is proportion to the negative log of the frequency Goal #2: Code should be unambiguous Since objects will be encoded at different lengths, as we read the bits, we need to know when we’ve reached the end of one object and should begin processing the next one This type of code is called a prefix code CMSC 100 -- Data Compression Tue 9/18/12 Using a Prefix Code How would you represent “HELLO” using this code? 0 0 0 1 A 0 H 1 1 E 1 L 0 1 O 0 9 Note: By convention, the left branch is 0; the right branch is 1 CMSC 100 -- Data Compression 1 C S Tue 9/18/12 Interpreting a Prefix Code 0 0 0 1 A 0 H 1 E 1 L What does “1110000110110111110” mean in this code? 10 CMSC 100 -- Data Compression 1 0 1 O 0 1 C S Tue 9/18/12 Interpreting a Prefix Code 0 0 0 1 A 0 H 11 1 E 1 L C What does “1110000110110111110” mean in this code? CMSC 100 -- Data Compression 1 0 1 O 0 1 C S Tue 9/18/12 Interpreting a Prefix Code 0 0 0 1 A 0 H 12 1 E 1 L C What does “1110 | 000110110111110” mean in this code? CMSC 100 -- Data Compression 1 0 1 O 0 1 C S Tue 9/18/12 Interpreting a Prefix Code 0 0 0 1 A 0 H 13 1 E 1 L C H What does “1110 | 000110110111110” mean in this code? CMSC 100 -- Data Compression 1 0 1 O 0 1 C S Tue 9/18/12 Interpreting a Prefix Code 0 0 0 1 A 0 H 14 1 L 1 1 E 0 1 O C H O O S E What does “1110 | 000 | 110 | 110 | 1111 | 10” mean in this code? CMSC 100 -- Data Compression 0 1 C S Tue 9/18/12 0 0 0 0 1 L 1 0 1 1 0 1 0 1 A SPC T 0 1 O 0 Y 0 1 1 0 W E 0 Decode the Message: 1 ! 1 C 0 1 0 1 M P S U 1 R 0111110010100101011011100011110111110110 010 00111111110 010 15 0110001110 010 0110001110 010 0110001110 010 0001100000100100000000110 010 011111001000000 01110 CMSC 100 -- Data Compression Tue 9/18/12 Encoding Algorithm Frequency distribution: Set of k objects, o1...ok Number of times of each object appears in the document, n1...nk Construct a Huffman code as follows: 1. Pick the two least frequent objects, oi and oj 2. Replace them with a single combined object, oij, with frequency ni+nj 3. If there are at least two objects left, go to step 1 Visually: Each of the original objects is a leaf (bottom node) in the prefix tree Each combined objects represents a 0/1 split where the “children” are the two objects that were combined In the last step, we combine two subtrees into a single final prefix tree 16 CMSC 100 -- Data Compression Tue 9/18/12 Encoding Example SHE SELLS SEASHELLS BY THE SEASHORE 17 CMSC 100 -- Data Compression Tue 9/18/12 Encoding Example SHE SELLS SEASHELLS BY THE SEASHORE Frequency distribution: 18 A–2 B–1 E–7 H–4 L–4 O–1 R–1 S–8 T–1 Y–1 <SPC> – 5 CMSC 100 -- Data Compression Tue 9/18/12 Encoding Example SHE SELLS SEASHELLS BY THE SEASHORE Frequency distribution: 19 A–2 B–1 E–7 H–4 L–4 O–1 R–1 S–8 T–1 Y–1 <SPC> – 5 CMSC 100 -- Data Compression 2 B1 O1 Tue 9/18/12 Encoding Example SHE SELLS SEASHELLS BY THE SEASHORE Frequency distribution: 20 A–2 B–1 E–7 H–4 L–4 O–1 R–1 S–8 T–1 Y–1 <SPC> – 5 CMSC 100 -- Data Compression 2 B1 2 O1 R1 3 T1 A2 Y1 Tue 9/18/12 Encoding Example SHE SELLS SEASHELLS BY THE SEASHORE Frequency distribution: 21 A–2 B–1 E–7 H–4 L–4 O–1 R–1 S–8 T–1 Y–1 <SPC> – 5 CMSC 100 -- Data Compression 7 4 2 B1 2 O1 R1 3 T1 A2 Y1 Tue 9/18/12 Encoding Example SHE SELLS SEASHELLS BY THE SEASHORE Frequency distribution: 22 A–2 B–1 E–7 H–4 L–4 O–1 R–1 S–8 T–1 Y–1 <SPC> – 5 CMSC 100 -- Data Compression 7 8 4 H4 2 B1 2 O1 R1 L4 3 T1 A2 Y1 Tue 9/18/12 Encoding Example SHE SELLS SEASHELLS BY THE SEASHORE Frequency distribution: 23 A–2 B–1 E–7 H–4 L–4 O–1 R–1 S–8 T–1 Y–1 <SPC> – 5 CMSC 100 -- Data Compression 12 _5 7 4 E7 H4 2 B1 8 2 O1 R1 L4 3 T1 A2 Y1 Tue 9/18/12 Encoding Example SHE SELLS SEASHELLS BY THE SEASHORE Frequency distribution: 24 A–2 B–1 E–7 H–4 L–4 O–1 R–1 S–8 T–1 Y–1 <SPC> – 5 CMSC 100 -- Data Compression 15 12 _5 7 4 E7 H4 2 B1 8 2 O1 R1 L4 3 T1 A2 Y1 Tue 9/18/12 Encoding Example SHE SELLS SEASHELLS BY THE SEASHORE 35 25 Frequency distribution: A–2 B–1 E–7 H–4 L–4 O–1 R–1 S–8 T–1 Y–1 <SPC> – 5 CMSC 100 -- Data Compression 20 15 12 _5 S8 8 4 E7 H4 2 B1 7 2 O1 R1 L4 3 T1 A2 Y1 Tue 9/18/12 Green Eggs and Ham 26 CMSC 100 -- Data Compression Tue 9/18/12 Green Eggs and Ham Symbols (not letters!) are words. Ignore spaces and punctuation. I am Sam I am Sam Sam I am 27 Do you like green eggs and ham? That Sam-I-am! I do not like them, That Sam-I-am! Sam-I-am. I do not like I do not like that Sam-I-am! CMSC 100 -- Data Compression green eggs and ham. Tue 9/18/12