Isaac's and marks log file Thursday Nov 9: Mark 10 hours work Isaac 8 hours work Organized. Found articles. Rewrote proposal. Planned work. Explored wavelets with dan in morning. Implemented DCT compound coder from K. and Tretter paper Saturday Nov 11: Isaac: 10 hours: Improved DCT Compund Quantizer. Made more general. Tracked bugs. Changed adaptive quantization parameters. Wrote findEntropy and findPSNR functions. Added min and max conditions on pixel and quantized table values, to solve a big bug. Changed the display of the image and the text blocks ("highlight"). Found images to use as examples. Sunday Nov 12 Isaac, Mark 7 hous Organized DCT Comp. Wrote notes.txt. Formalized what variables and measures we would vary and collect. Wrote down modifcations that we planned to implement. Implemented Region PSNR and Entropy measures, smoothing filters for activity matrix, uniform DC quatization within the compound coder. Tuesday Nov 15 Isaac 3 hours Read main articles on adaptive image coding and text segmentation in mixed-mode images to check the work so far. Analyzed main ways to compute the activity of a block, and main ways we should use. Compared results of those ways as explained in papers. Looked for new articles on same subject. Wednesday Nov 16 Mark, Isaac, 8.5 hours Tried to come up with a simple and quick alternative to entropy calculation for a rate metric. Discussed a fixed-length run-length code with Markus. Implemented, tested and debugged various versions of this rate metric. Looked for more images to code. Tried to find images that were good examples. Thursday Nov 17 Morning Mark 5 hours, Isaac, 4 hours. Developed histogram-based entropy calculations that finally makes mark comfortable. 5 images each were found as training sets for text, graph, and compound. Histograms were made for each class of image, for 10 different quantization scales, and for each of the 64 DCT coefficients. Wrote new functions in transcode.m to calculate entropies using these histograms. Evening Mark 9:30 [40], Isaac 9:30, [50] Collected compound images from stanford online. Made PSNR vs Rate curves for text and for graphs. Manually segmented compound images. Implemented a new DCT coefficient-based segmentation algorithm. Implemented text locator scoring with false positives negatives and errors. Wednesday Nov. 22 Mark 7, Isaac 8, Isaac implemented a couple more text location schemes. Mark and Isaac discussed how the schemes should be compared. Decided to use two images to find the optimal threshold setting (as a training set) for each scheme, and then use three image to compare the performance of the schemes for that threshold. Mark began writing a script to do all the comparisons of the various schemes. Thursday Nov 23 Happy Thanksgiving! Mark 6.5 Isaac 1 Mark finished writing the script "CompareSchemes3.m" and collected graphs, optimal threshold values. This was done for equally weighted falseNegatives and FalsePositives and also for falseNegatives given twice the cost of falsePositives. Isaac came in and tweaked the threshold ranges and advised. Then we went to Safeway to get wine and pies for thanksgiving dinners. Mark intends to start writing the report tonight. Friday: Isaac: 12, Mark: 13 [71/65] Isaac: Read many papers and implemented many schemes from these papers, including edge based, variance based, abs. Deviation based, energy based, entropy based, and an innovative one, 1st BIG improvement, histogram-based. Ran very slow program to compare all the schemes. Mark: Tweaked Compare schemes 3. Rate Psnr curves. Started working on smoothing. Made a lot of different plots. Fixed PSNR functions and entropy fncion to run faster. Region based entropies and PSNR using 'true Activities' Mark, Isaac:Made the outline and lists of what needed to get done. Mark Saturday: Mark: 11 Isaac: 11 Mark: Implemented Non Linear Smoothing. Evaluated its performance with CompareSchemes3. Collected graphs for the Girod example. Isaac: thought of new schemes based on the innovative one, that should be useful in the other parts of the report - such as high-frequency based (horizontal, vertical, both), some coefficients..., with Mark 2nd BIG improvement: a MAP-rule (Maximum A-Posteori) based approach., inverse energy, inverse text, only graph , and a 3rd BIG improvement - basing it on non quantized DCT histogram (only next integer rounded)