Recherche avancée

Sur d’autres sites (4622)

• Revision 28933 : on bouge

  31 mai 2009, par ben.spip@… — Log

  on bouge

• Developing MobyCAIRO

  26 mai 2021, par Multimedia Mike — General

  I recently published a tool called MobyCAIRO. The ‘CAIRO’ part stands for Computer-Assisted Image ROtation, while the ‘Moby’ prefix refers to its role in helping process artifact image scans to submit to the MobyGames database. The tool is meant to provide an accelerated workflow for rotating and cropping image scans. It works on both Windows and Linux. Hopefully, it can solve similar workflow problems for other people.

  As of this writing, MobyCAIRO has not been tested on Mac OS X yet– I expect some issues there that should be easily solvable if someone cares to test it.

  The rest of this post describes my motivations and how I arrived at the solution.

  Background
  I have scanned well in excess of 2100 images for MobyGames and other purposes in the past 16 years or so. The workflow looks like this :

  
  

  

  Image workflow

  
  

  It should be noted that my original workflow featured me manually rotating the artifact on the scanner bed in order to ensure straightness, because I guess I thought that rotate functions in image editing programs constituted dark, unholy magic or something. So my workflow used to be even more arduous :

  
  

  

  I can’t believe I had the patience to do this for hundreds of scans

  
  

  Sometime last year, I was sitting down to perform some more scanning and found myself dreading the oncoming tedium of straightening and cropping the images. This prompted a pivotal question :

  
  Why can’t a computer do this for me ?
  

  After all, I have always been a huge proponent of making computers handle the most tedious, repetitive, mind-numbing, and error-prone tasks. So I did some web searching to find if there were any solutions that dealt with this. I also consulted with some like-minded folks who have to cope with the same tedious workflow.

  I came up empty-handed. So I endeavored to develop my own solution.

  Problem Statement and Prior Work
  
  I want to develop a workflow that can automatically rotate an image so that it is straight, and also find the most likely crop rectangle, uniformly whitening the area outside of the crop area (in the case of circles).

  As mentioned, I checked to see if any other programs can handle this, starting with my usual workhorse, Photoshop Elements. But I can’t expect the trimmed down version to do everything. I tried to find out if its big brother could handle the task, but couldn’t find a definitive answer on that. Nor could I find any other tools that seem to take an interest in optimizing this particular workflow.

  When I brought this up to some peers, I received some suggestions, including an idea that the venerable GIMP had a feature like this, but I could not find any evidence. Further, I would get responses of “Program XYZ can do image rotation and cropping.” I had to tamp down on the snark to avoid saying “Wow ! An image editor that can perform rotation AND cropping ? What a game-changer !” Rotation and cropping features are table stakes for any halfway competent image editor for the last 25 or so years at least. I am hoping to find or create a program which can lend a bit of programmatic assistance to the task.

  Why can’t other programs handle this ? The answer seems fairly obvious : Image editing tools are general tools and I want a highly customized workflow. It’s not reasonable to expect a turnkey solution to do this.

  Brainstorming An Approach
  I started with the happiest of happy cases— A disc that needed archiving (a marketing/press assets CD-ROM from a video game company, contents described here) which appeared to have some pretty clear straight lines :

  
  
  

  My idea was to try to find straight lines in the image and then rotate the image so that the image is parallel to the horizontal based on the longest single straight line detected.

  I just needed to figure out how to find a straight line inside of an image. Fortunately, I quickly learned that this is very much a solved problem thanks to something called the Hough transform. As a bonus, I read that this is also the tool I would want to use for finding circles, when I got to that part. The nice thing about knowing the formal algorithm to use is being able to find efficient, optimized libraries which already implement it.

  Early Prototype
  A little searching for how to perform a Hough transform in Python led me first to scikit. I was able to rapidly produce a prototype that did some basic image processing. However, running the Hough transform directly on the image and rotating according to the longest line segment discovered turned out not to yield expected results.

  
  
  

  It also took a very long time to chew on the 3300×3300 raw image– certainly longer than I care to wait for an accelerated workflow concept. The key, however, is that you are apparently not supposed to run the Hough transform on a raw image– you need to compute the edges first, and then attempt to determine which edges are ‘straight’. The recommended algorithm for this step is the Canny edge detector. After applying this, I get the expected rotation :

  
  
  

  The algorithm also completes in a few seconds. So this is a good early result and I was feeling pretty confident. But, again– happiest of happy cases. I should also mention at this point that I had originally envisioned a tool that I would simply run against a scanned image and it would automatically/magically make the image straight, followed by a perfect crop.

  Along came my MobyGames comrade Foxhack to disabuse me of the hope of ever developing a fully automated tool. Just try and find a usefully long straight line in this :

  
  

  Darn it, Foxhack…
  

  There are straight edges, to be sure. But my initial brainstorm of rotating according to the longest straight edge looks infeasible. Further, it’s at this point that we start brainstorming that perhaps we could match on ratings badges such as the standard ESRB badges omnipresent on U.S. video games. This gets into feature detection and complicates things.

  This Needs To Be Interactive
  At this point in the effort, I came to terms with the fact that the solution will need to have some element of interactivity. I will also need to get out of my safe Linux haven and figure out how to develop this on a Windows desktop, something I am not experienced with.

  I initially dreamed up an impressive beast of a program written in C++ that leverages Windows desktop GUI frameworks, OpenGL for display and real-time rotation, GPU acceleration for image analysis and processing tricks, and some novel input concepts. I thought GPU acceleration would be crucial since I have a fairly good GPU on my main Windows desktop and I hear that these things are pretty good at image processing.

  I created a list of prototyping tasks on a Trello board and made a decent amount of headway on prototyping all the various pieces that I would need to tie together in order to make this a reality. But it was ultimately slowgoing when you can only grab an hour or 2 here and there to try to get anything done.

  Settling On A Solution
  Recently, I was determined to get a set of old shareware discs archived. I ripped the data a year ago but I was blocked on the scanning task because I knew that would also involve tedious straightening and cropping. So I finally got all the scans done, which was reasonably quick. But I was determined to not manually post-process them.

  This was fairly recent, but I can’t quite recall how I managed to come across the OpenCV library and its Python bindings. OpenCV is an amazing library that provides a significant toolbox for performing image processing tasks. Not only that, it provides “just enough” UI primitives to be able to quickly create a basic GUI for your program, including image display via multiple windows, buttons, and keyboard/mouse input. Furthermore, OpenCV seems to be plenty fast enough to do everything I need in real time, just with (accelerated where appropriate) CPU processing.

  So I went to work porting the ideas from the simple standalone Python/scikit tool. I thought of a refinement to the straight line detector– instead of just finding the longest straight edge, it creates a histogram of 360 rotation angles, and builds a list of lines corresponding to each angle. Then it sorts the angles by cumulative line length and allows the user to iterate through this list, which will hopefully provide the most likely straightened angle up front. Further, the tool allows making fine adjustments by 1/10 of an angle via the keyboard, not the mouse. It does all this while highlighting in red the straight line segments that are parallel to the horizontal axis, per the current candidate angle.

  
  
  

  The tool draws a light-colored grid over the frame to aid the user in visually verifying the straightness of the image. Further, the program has a mode that allows the user to see the algorithm’s detected edges :

  
  
  

  For the cropping phase, the program uses the Hough circle transform in a similar manner, finding the most likely circles (if the image to be processed is supposed to be a circle) and allowing the user to cycle among them while making precise adjustments via the keyboard, again, rather than the mouse.

  
  
  

  Running the Hough circle transform is a significantly more intensive operation than the line transform. When I ran it on a full 3300×3300 image, it ran for a long time. I didn’t let it run longer than a minute before forcibly ending the program. Is this approach unworkable ? Not quite– It turns out that the transform is just as effective when shrinking the image to 400×400, and completes in under 2 seconds on my Core i5 CPU.

  For rectangular cropping, I just settled on using OpenCV’s built-in region-of-interest (ROI) facility. I tried to intelligently find the best candidate rectangle and allow fine adjustments via the keyboard, but I wasn’t having much success, so I took a path of lesser resistance.

  Packaging and Residual Weirdness
  I realized that this tool would be more useful to a broader Windows-using base of digital preservationists if they didn’t have to install Python, establish a virtual environment, and install the prerequisite dependencies. Thus, I made the effort to figure out how to wrap the entire thing up into a monolithic Windows EXE binary. It is available from the project’s Github release page (another thing I figured out for the sake of this project !).

  The binary is pretty heavy, weighing in at a bit over 50 megabytes. You might advise using compression– it IS compressed ! Before I figured out the --onefile command for pyinstaller.exe, the generated dist/ subdirectory was 150 MB. Among other things, there’s a 30 MB FORTRAN BLAS library packaged in !

  Conclusion and Future Directions
  Once I got it all working with a simple tkinter UI up front in order to select between circle and rectangle crop modes, I unleashed the tool on 60 or so scans in bulk, using the Windows forfiles command (another learning experience). I didn’t put a clock on the effort, but it felt faster. Of course, I was livid with proudness the whole time because I was using my own tool. I just wish I had thought of it sooner. But, really, with 2100+ scans under my belt, I’m just getting started– I literally have thousands more artifacts to scan for preservation.

  The tool isn’t perfect, of course. Just tonight, I threw another scan at MobyCAIRO. Just go ahead and try to find straight lines in this specimen :

  
  
  

  I eventually had to use the text left and right of center to line up against the grid with the manual keyboard adjustments. Still, I’m impressed by how these computer vision algorithms can see patterns I can’t, highlighting lines I never would have guessed at.

  I’m eager to play with OpenCV some more, particularly the video processing functions, perhaps even some GPU-accelerated versions.

  The post Developing MobyCAIRO first appeared on Breaking Eggs And Making Omelettes.

• Subtitling Sierra VMD Files

  1er juin 2016, par Multimedia Mike — Game Hacking

  I was contacted by a game translation hobbyist from Spain (henceforth known as The Translator). He had set his sights on Sierra’s 7-CD Phantasmagoria. This mammoth game was driven by a lot of FMV files and animations that have speech. These require language translation in the form of video subtitling. He’s lucky that he found possibly the one person on the whole internet who has just the right combination of skill, time, and interest to pull this off. And why would I care about helping ? I guess I share a certain camaraderie with game hackers. Don’t act so surprised. You know what kind of stuff I like to work on.

  The FMV format used in this game is VMD, which makes an appearance in numerous Sierra titles. FFmpeg already supports decoding this format. FFmpeg also supports subtitling video. So, ideally, all that’s necessary to support this goal is to add a muxer for the VMD format which can encode raw video and audio, which the format supports. Implement video compression as extra credit.

  The pipeline that I envisioned looks like this :

  
  

  

  VMD Subtitling Process

  
  

  “Trivial !” I surmised. I just never learn, do I ?

  The Plan
  So here’s my initial pitch, outlining the work I estimated that I would need to do towards the stated goal :

  1. Create a new file muxer that produces a syntactically valid VMD file with bogus video and audio data. Make sure it works with both FFmpeg’s playback system as well as the proper Phantasmagoria engine.
  2. Create a new video encoder that essentially operates in pass-through mode while correctly building a palette.
  3. Create a new basic encoder for the video frames.

  A big unknown for me was exactly how subtitle handling operates in FFmpeg. Thanks to this project, I now know. I was concerned because I was pretty sure that font rendering entails anti-aliasing which bodes poorly for keeping the palette count under 256 unique colors.

  Computer Science Puzzle
  When pondering how to process the palette, I was excited for the opportunity to exercise actual computer science. FFmpeg converts frames from paletted frames to full RGB frames. Then it needs to convert them back to paletted frames. I had a vague recollection of solving this problem once before when I was experimenting with a new paletted video codec. I seem to recall that I did the palette conversion in a very naive manner. I just used a static 256-element array and processed each RGB pixel of the frame, seeing if the value already occurred in the table (O(n) lookup) and adding it otherwise.
  

  There are more efficient algorithms, however, such as hash tables and trees. Somewhere along the line, FFmpeg helpfully acquired a rarely-used tree data structure, which was perfect for this project.

  So I was pretty pleased with this optimization. Too bad this wouldn’t survive to the end of the effort.

  Another palette-related challenge was the fact that a group of pictures would be accumulating a new palette but that palette needed to be recorded before the group. Thus, the muxer needed to have extra logic to rewind the file when the video encoder transmitted a palette change.

  Video Compression
  VMD has a few methods in its compression toolbox. It can use interframe differencing, it has some RLE, or it can code a frame raw. It can also use a custom LZ-like format on top of these. For early prototypes, I elected to leave each frame coded raw. After the concept was proved, I implemented the frame differencing.

  
  

  

  
  Top frame compared with the middle frame yields the bottom frame : red pixels indicate changes
  

  Encoding only those red dots in between vast runs of unchanged pixels yielded a vast measurable improvement. The next step was to try wiring up FFmpeg’s existing LZ compression facilities to the encoder. This turned out to be implausible since VMD’s LZ variant has nothing to do with anything FFmpeg already provides. Fortunately, the LZ piece is not absolutely required and the frame differencing + RLE provides plenty of compression.

  Subtitling
  I’ve never done anything, multimedia programming-wise, concerning subtitles. I guess all the entertainment I care about has always been in my native tongue. What a good excuse to program outside of my comfort zone !

  First, I needed to know how to access FFmpeg’s subtitling facilities. Fortunately, The Translator did the legwork on this matter so I didn’t have to figure it out.

  However, I intuitively had misgivings about this phase. I had heard that the subtitling process performs anti-aliasing. That means that the image would need to be promoted to a higher colorspace for this phase and that the anti-aliasing process would likely push the color count way past 256. Some quick tests revealed this to be the case, as the running color count would leap by several hundred colors as soon as the palette accounting algorithm encountered a subtitle.

  So I dug into the subtitle subsystem. I discovered that the subtitle library operates by creating a linked list of subtitle bitmaps that the client app must render. The bitmaps are comprised of 8-bit alpha transparency values that must be composited onto the target frame (i.e., 0 = transparent, 255 = 100% opaque). For example, the letter ‘H’ :

                                    (with 00s removed)
  13 F8 41 00 00 00 00 68 E4  |  13 F8 41             68 E4    
  14 FF 44 00 00 00 00 6C EC  |  14 FF 44             6C EC
  14 FF 44 00 00 00 00 6C EC  |  14 FF 44             6C EC
  14 FF 44 00 00 00 00 6C EC  |  14 FF 44             6C EC
  14 FF DC D0 D0 D0 D0 E4 EC  |  14 FF DC D0 D0 D0 D0 E4 EC
  14 FF 7E 50 50 50 50 9A EC  |  14 FF 7E 50 50 50 50 9A EC
  14 FF 44 00 00 00 00 6C EC  |  14 FF 44             6C EC
  14 FF 44 00 00 00 00 6C EC  |  14 FF 44             6C EC
  14 FF 44 00 00 00 00 6C EC  |  14 FF 44             6C EC
  11 E0 3B 00 00 00 00 5E CE  |  11 E0 3B             5E CE
  

  To get around the color explosion problem, I chose a threshold value and quantized values above and below to 255 and 0, respectively. Further, the process chooses an appropriate color from the existing palette rather than introducing any new colors.

  Muxing Matters
  In order to force VMD into a general purpose media framework, a lot of special information needs to be passed around. Like many paletted codecs, the palette needs to be transmitted from the file demuxer to the video decoder via some side channel. For re-encoding, this also implies that the palette needs to make the trip from the video encoder to the file muxer. As if this wasn’t enough, individual VMD frames have even more data that needs to be ferried between the muxer and codec levels, including frame change boundaries. FFmpeg provides methods to do these things, but I could not always rely on the systems to relay the data in all cases. I was probably doing something wrong ; I accept that. Instead, I just packed all the information at the front of an encoded frame and split it apart in the muxer.

  I could not quite figure out how to get the audio and video muxed correctly. As a result, neither FFmpeg nor the Phantasmagoria engine could replay the files correctly.

  Plan B
  Since I was having so much trouble creating an entirely new VMD file, likely due to numerous unknown bits of the file format, I thought of another angle : re-use the existing VMD file. For this approach, I kept the video encoder and file muxer that I created in the initial phase, but modified the file muxer to emit a special intermediate file. Then, I created a Python tool to repackage the original VMD file using compressed video data in the intermediate file.

  For this phase, I also implemented a command line switch for FFmpeg to disable subtitle blending, to make the feature feel like less of an unofficial hack, as though this nonsense would ever have a chance of being incorporated upstream.

  At this point, I was seeing some success with the complete, albeit roundabout, subtitling process. I constructed a subtitle file using “Spanish I Learned From Mexican Telenovelas” and the frames turned out fairly readable :

  
  

  

  “she cheated on him”

  
  

  

  “he’s a scumbag” … these random subtitles could fit surprisingly well !

  
  

  The few files that I tested appeared to work fine. But then I handed off my work to The Translator and he immediately found a bunch of problems. According to my notes, the problems mostly took the form of flashing, solid color frames. Further, I found tiny, mostly imperceptible flaws in my RLE compressor, usually only detectable by running strict comparison tools ; but I wasn’t satisfied.

  At this point, I think I attempted to just encode the entire palette at the front of each frame, as allowed by the format, but that did not seem to fix any problems. My notes are not completely clear on this matter (likely because I was still trying to figure out the exact problem), but I think it had to do with FFmpeg inserting extra video frames in order to even out gaps in the video framerate.

  Sigh, Plan C
  At this point, I was getting tired of trying to force FFmpeg to do this. So I decided to minimize its involvement using lessons learned up to this point.

  The next pitch :

  1. Create a new C program that can open an existing VMD file and output an identical VMD file. I know this sounds easy, but the specific method of copying entails interpreting individual parts of the file and writing those individual parts to the new file. This is in preparation for…
  2. Import the VMD video decoder functions directly into the program to decode the individual video frames and re-encode them, replacing the video frames as the file is rewritten.
  3. Wire up the subtitle system. During the adventure to disable subtitle blending, I accidentally learned enough about interfacing to the subtitle library to just invoke it directly.
  4. Rewrite the RLE method so that it is 100% correct.

  Off to work I went. That part about lifting the existing VMD decoder functions out of their libavcodec nest turned out to not be that straightforward. As an alternative, I modified the decoder to dump the raw frames to an intermediate file. In doing so, I think I was able to avoid the issue of the duplicated frames that plagued the previous efforts.

  Also, remember how I was really pleased with the palette conversion technique in which I was able to leverage computer science big-O theory ? By this stage, I had no reason to convert the paletted video to RGB in the first place ; all of the decoding, subtitling and re-encoding operates in the paletted colorspace.

  This approach seemed to work pretty well. The final program is subtitle-vmd.c. The process is still a little weird. The modifications in my own FFmpeg fork are necessary to create an intermediate file that the new C tool can operate with.

  Next Steps
  The Translator has found some assorted bugs and corner cases that still need to be ironed out. Further, for extra credit, I need find the change windows for each frame to improve compression just a little more. I don’t think I will be trying for LZ compression, though.

  However, almost as soon as I had this whole system working, The Translator informed me that there is another, different movie format in play in the Phantasmagoria engine called ROBOT, with an extension of RBT. Fortunately, enough of the algorithms have been reverse engineered and re-implemented in ScummVM that I was able to sort out enough details for another subtitling project. That will be the subject of a future post.

  See Also :

  • Subtitling Sierra RBT Files : The followup in which I discuss how to scribble text on the other animation format

  The post Subtitling Sierra VMD Files first appeared on Breaking Eggs And Making Omelettes.