FPT Toolkit and my first steps into win32

 The screenshot above is my latest little project, a windows based tool for the decomposition of image using my fast pulse transform (FPT) technique. It needed to have a few features like file IO, pulse selections and highlighting of pulses. I’ve never done any sort of c++ winapi programming before apart from some windows forms stuff in c#. I figured since i’m busy messing around with win32 in my openGL and directX adventures, i might as well just use that. :/

Probably not the smartest decision, since the bare win api is c based and not c++ based and so the way the api is structured is based on c techniques. I’ll be honest and it took me a day or two to get used to it and to figure out how most things occurs and now i’ve pretty much got the gist of it. Now it comes down to using the msdn docs to find out what the flag values i need are, and trying to figure out when specific messages are sent. I’ve got a very basic grasp of it at the moment. I wish i could spend more time messing around with it but i need to focus on this semester and finish learning directx 10. And even so once that is done, i think i’m gonna focus on learning MFC or windows forms even as it seems the bulk of development is done using these APIs. From what i can see most window game dev toolsets are developed in MFC.

So some last words about my app… i couldn’t for the life of me figure out how the fuck DIBs and HBITMAPS work (creating DIBs programmatically at least – the conversion from DIB to HBITMAP was simple enough) so i had to use openGL to display my 2D images (haha, oh well its not like people don’t have openGL videos cards these days). I just got the pulse selection and reconstruction saving working, just need to finish the highlighting section and its done (for now). I need to sit and think logically about what i want for the highlighting interface and then code it.

I have struggled these last few days trying to learn winapi programming from scratch using nothing more than google and the MSDN docs but at the same time I’ve learnt a lot and so its been worth it! OH and for a development environment: visual studio 2008 + vista 64 is a dream come true

Radon Transform C++ Implementation Update

Thanks to haDoan for noticing that i had left out the definition for the pixel struct in the source code. When i put up the code, i ripped it out of my surveillance project and didnt noticed the definition was located in a different file.

I’ve updated the source code to include the struct definition : http://www.cs.up.ac.za/cs/banguelov/blog/radonTransformer.h

DPT Performance and Stack Overflows

I finally finished the private contract I was doing on saturday, and so I sat down sunday morning to start generating results for the presentation I am going to be doing for the CSIR regarding this on thursday. Things went well until about 2 minutes in when the program crashed, running it from within VS, it dies with an unhandled exception: stack overflow… I panicked and instead of driving to work to do some research on stack overflows, I assumed that my code was broken. After debugging for around 8 hours with no results apart from figuring out that the problem was in the grow operation, I had a mini breakdown. I left it all and dedicated the entire day today to fixing my “broken” code. It was the best thing I could have done!

I came in this morning feeling calm and relaxed, sat down and ran my program again and realized that it was an exception being thrown and not a fatal error! Did a google search on the exception and figured out it was due to my program stack being too small, so my “broken” code wasnt broken at all but rather it was suffering from a severe case of PEBCAK (google it )…

The locate critical sets function makes use of recursion to locate the initial sets, and on outside images there are large areas (sky) that are the same color so it starting recursion thousands of levels deep and obviously the default 1mb stack cant store all that, increasing the stack size to 25mb fixed everything! If anyone is interested the option for setting the stack size is found under project -> properties -> linker -> System -> Stack Reserve Size (in bytes).

 So I ran the simulation and the results are interesting, i ran it on 10 different images, with 9 different sizes ranging for 100% to 10% in 10% intervals. Apart from a few anomalies the graphs look pretty linear, this is alot better than the previous results where the results seem exponential. See for yourself.

UPDATE: These are the updated performance graphs from the averaged results from 10 runs. I’m still not too sure what to make of them yet so i wont even hazard a guess… One thing is clear tho that we need to do more research…

Figure 1. Time taken to locate initial Critical Sets.

Figure 2. Time taken to decompose the image.

DPT Code and Pseudocode

okay here’s the code and the pseudocode of the DPT implementation, any comments on style, performance and so on are more than welcome…

 It’s been a few weeks of work to get this done, the algorithm has changed significantly since i first received, from speed optimizations to the removal of redundant code, merging of function to reduce complexity and so on. I for one am happy with the current state.

matlab_src.zip – original matlab source, if you’re interested in the before

pseudocode.zip  – basic pseudocode overview of current implementation

DPT_initial.zip - c++ implementation and visual studio project file

DPT initial implementation complete!

OH MY GOD! its finally done! You have no idea how happy i am right now… (and pretty drained)

 I’ve spent the last three days debugging the conversion i did from matlab to c++. The problem was that the program would complete but the image wouldnt have been totally cleared (ie. be a fixed value throughout). Okay i’m getting ahead of myself. Basically the DPT algorithm (which we still have to name) seperates an image according to the contrast grouping. Its basically an intense smoother.

 It starts out by identifying all the local minimum and maximum pixel groupings, it does this by recursively finding all pixel groupings and each groupings adjacent pixels (border pixels), it then compares the pixel groups color value to its adjacent pixels, if it is larger or smaller than all these values, then the group is a maximum or minimum respectively.

This seems to be the most intensive operation in the algorithm , i havent had time to slap it through a profiler yet so i cant be sure of this.

Once this initial “search” is complete, we start extracting “pulses”/pixel groups of a certain size. We start at size 1 and continue until there is only one group left (ie, the entire image has been smoothed down to a single color). When we extract the group, we check if it is a max or a min group, if so then we set the value of the pixels in that group to the max or min value of the adjacent pixels, and then we “grow” to see if it will now expand to the neighboring pixel, it may even overlap with another grouping, if this occurs we simply merge the two groups. By setting the groups pixel to a value neighbouring it we’re always guaranteed that the group will grow, so we are always guaranteed to full smooth an image.

Max groups are called “up pulses” and min groups are “down pulses”, once the image has been cleared we can then use these extracted “pulses” and reconstruct the image from the pulses, either full or partially. I’ve provided sample output of the program below, i entered in an image and then did several partial reconstructions with the extracted pulses.

Original Image

Partial Reconstruction: Grouping size 200 – 350

Partial Reconstruction: Grouping size 350 – 800

Partial Reconstruction: Grouping size 1000 – 2000

Partial Reconstruction: Grouping size 2000 – 35000

Partial Reconstruction: Grouping size 35000 – 60000

Partial Reconstruction: Grouping size 35000 – 85000

Kinda cool huh? Anyways the really funny thing is now that the algorithm is complete, we need to figure out the application for the algorithm. Kinda like developing a cure for a disease that hasnt been discovered yet. Should prove fun…

The next step for me know is to develop a win32 application that will function as a work bench for this algorithm so that we can easily see the results, i honestly have no idea how complicated it will be to embed my ANSI c++ code into a visual c++ program and so that will be my next challenge. Keep an eye out.

Also since this is now complete (i was envisioning debugging code for most of the weekend) I can now spend a large chunk of the weekend working on the design document for our mod. That’s gonna be a lot of work but work that i’m looking forward to.

*bobby is now off to shower and collapse

Radon Transform

Wow, it’s been a rough couple of weeks: I had to hand in my graphics project, study for a statistics test, fighting off my allergies (I hate spring) and then I had to study for my finals. At least I have my degree now.

Finally!  

Anyways, I promised I’d write about the radon transformation I used to convert from the extracted images to a numerical format suitable for input into our neural network. This technique is extremely effective and is already used in industry for just such purposes.  We tested it on demo day with very minimal data and it worked remarkably well.

Before I get knee deep in the technical aspects of the system, I need to mention this: due to the preprocessing done on the motion detected, there is no need for a complicated AI system; the radon transformation and the recursive feature extractor together remove a lot of noise and problems that may have been present otherwise. The radon transform especially helps as we have built in scaling so this does not have to be taken into account later on. Also from the results of the transformation, objects similar in shape have extremely similar radon transformations so the training time of the neural network was reduced as was the amount of hidden neurons necessary.

In the final demo we used a neural network with 408 inputs and only 4 hidden neurons, scary isn’t it.

Introduction:

Now back to the nitty gritty: the radon transform. If you Google the “radon transform” you’ll probably get the Wikipedia page with a scary looking equation.  I also got a fright the first time I saw this but after some research it’s really simple.

The basic idea of the radon transform (or my modified version thereof) is simple: if you look at your 2D image in the XY plane, you simply flatten the image onto the X axis (figure1), then divide the X axis into several beams and you work out the amount of pixels within each beam. Your output will be the pixel contributions of the object to each beam.  Then you’d rotate the object and flatten it once again. Doing this for multiple angles will give you a very good representation of the objects shape.

Figure 1

The most basic (and unfortunately most commonly used technique) for image classification is to simple get the centroid of the object and then trace the outline of the object giving you a silhouette. Now this doesn’t sound so bad does it? Well, it is firstly it doesn’t handle broken up images well (not without major preprocessing or modification) and it also loses a lot of detail and can provide false matches. In figure1 below we have the radon transform of a solid circle and a hollow circle, a standard outline trace would provide the exact shape result for these obviously different shapes while as you can see the radon transform (in one projection) provides completely different results.  Again this pre-processing will take the strain of the neural network (or other AI technique we’ll use for classification).

Figure 2

Okay now for the technical details: as you remember we flatten the image according to some projection. Figure2 shows some of these projections. Now if you look at figure 2 you might notice that the now flattened image’s top border can be seen as a graph of some function, so the amount of pixels in a beam is the approximate area under the graph between the left and right end points of the beam.

Now that picture is misleading as you might think that that it is a square object that we’ve rotated and flattened, but it is in fact a single pixel. The algorithm works on a per pixel basis. Instead of actually flattening the object, we simply work out the equation of the graph for a single rotated pixel and then use that to run through all the pixels in the object, work out the left most and right most and then add them to their respective beams.  

Now some of you are screaming that if we just rotate the pixels it will be wrong as we aren’t rotating the entire object but that is taken into account later on.

Now how do we calculate the area under the graph for each pixel and how do we figure out what beam to add it to since a beam will have lots and lots of pixels in it? Also the beam widths will differ per object.

Figure 3

What we do is simply divide the beams into lots of sub-beams, so that multiple sub-beams pass through each pixel. Then for each pixel we work out the left most sub-beam and the right most sub-beam that passes through the pixel. This then becomes the domain for the equation of the graph we have earlier and we loop through each sub-beam, calculate the pixels contribution to it(the area under the graph) and then add it to the sub-beam total. This is shown in figure 3. What you also notice from figure 3 is that there is a small degree of approximation to reduce the calculations required for the area, but remember that we’re talking about fractions of a pixel here so the total error in approximation can easily be ignored.

Now for each projection we run through each pixel and add it to the appropriate sub-beam. Once this is complete we sum the sub-beams up into the initial amount of beams and then we divide each beam by the scaling factor. The scaling factor is simply the total pixels over the beam width; this reduces the total area for the beams to 1. So every object gets reduces to an n-beam representation where the sum of all beams is equal to 1.

Okay, my explanation is very basic and I’m sure mathematicians would point out various mistakes and so on , but I’m trying to make this easy to understand and to follow, it is not meant as a 100% mathematically accurate explanation, obviously if you wish to implement something like this, you wouldn’t only use my guide here as a reference. I’ve also left out some details but they should become apparent from the below explanations.

I’m struggling to find a good way to structure this guide so I’m just going to run through the algorithm simply just to finish off. 

Preparation:

The first steps we need to take before we can process the object is to get the total number of pixels, work out the centroid and the approximate radius of the object. Using the radius we work out the amount of beams and sub-beams we need for the transformation. Remember that we want several sub-beams to pass through each pixel.

Projection:

Now we run each of our projection functions to calculate the sub-beams totals. I’ll run through the basic procedure for a  projection:Work out the center of the pixel on the new axis (this where the rotation of the object comes into play)Work out the left most and right most sub-beams that pass through the pixel.For each sub-beam add the pixel’s contribution to it.

Note: For some equations there is an incline to the graph and so this needs to be calculated too, and processed separately. I.e. Work out the left most and right most sub-beams for the increasing incline and the decreasing incline and then using that work out out all the section separately.

Combine Sub-beams:

Now once all the sub-beams have been calculated, we work out the scaling factor which is: beamWidth /numPixels. We then sum all the sub-beams into beams (per projection) and multiply each one by the scaling factor. And that’s it. We have our complete numerical representation of our image.

Note: I used only 8 projections as I had very limited CPU time left at this stage of the project and had to limit the amount of processing that needs to be done, obviously more projections will be better but then again too many would be worse. A fine balance needs to be found, I personally think that 8 projections are more than sufficient for my needs. Again GPGPU programming would be so useful here!

C++ Source Code: radonTransformer.h

Feature Extractor – Initial Results

Okay, it’s been a while since the last update. I coded the outline extractor as I discussed in my previous post. It works for small objects but once most of the screen or lots of little objects occur it crashes. I think its a memory allocation problem related to my use of STL vectors. Anyways It turn out that it might not be the best method to use. Especially due to the image reconstruction problem, if I just get a list of little items (possibly hundreds of them) doing a compare between all of them will be extremely expensive and not too mention quite complex as a simple vertical/horizontal distance compare wont work.

What I am now using is a recursive extraction method that extracts the entire object even if it is broken up by using a neighborhood check in each step. Simple put it searches for a white pixel in the motion frame, once it is found it creates a new object and adds the pixel’s position to the object. Then it finds that pixels neighbors which are white and adds them and for each of those neighbors it finds their neighbors that are white and adds them and so on. The trick in this algorithm is in the finding of the neighbors…

This is done by checking all the pixels around a pixel within a radius of r pixels. so for a radius of 1, it will check the 8 surrounding pixels ( the neighborhood if you will ) . A radius of 2 will increase this neighborhood to 24 pixels, 3 will make it 48 pixels and so on. By modifying this radius I can now extract whole objects which are made up of lots of closely spaced sections. The below image illustrates this technique, note that the blue dot is the starting point.

This technique extracts multiple objects reasonably quickly assuming that they are not too large. The cost of the algorithm increases dramatically when the number of separate objects is increased or the size of each object is increased. It also randomly crashes when the entire screen is white, I’m assuming it’s a memory allocation error and will look into it.

Pseudo code for the algorithm is as follows (assuming you’ve already found an initial white pixel):

using radius:
    calculate range (rowStart,rowEnd,colStart,colEnd) for checking   

for ( r = rowStart, r < rowEnd, r++ )
{ 
    for ( c = colStart, c < colEnd, c++ )
    {
        if ( pixel(r,c) = white )
        {
            add it to the object
            set pixel to gray
        }
    }
}     

if (neighbors are found)
{
    foreach ( neighbor found)
    {
        //recursion occurs here
        add its neighbors to the object
    }
}
else
{
    return empty neighbor list
}

I tried optimizing this entire procedure by avoiding the calculation of the neighborhood on each step by simply creating a massive lookup table with each position and its neighbors and just looking it up instead of calculating it but much to my surprise once i profiled the code, it was significantly slower to lookup the neighbors than to calculate them?! I have a feeling its got something to do with internal copying of vectors when I return the neighbors. I guess this weekend is gonna be spent staring at a profiler again.

Now once the extractor is complete, I have to transform the object into a suitable format to input into my neural network, i was thinking of the centroid method i described in my previous post but my dad feels it would be much better to run the object through a radon transform and using the resulting data for the neural network. I’m still a bit away from that right now but once i get there I’ll let you know know what i find out.

Update: c++ source code – recursiveExtractor.cpp

Feature Extractor Intro

I’m currently busy working on a feature extractor for my final year project. I’m doing an outdoor surveillance system with shape classification. Currently I’m using a sigma delta background estimation algorithm (removes areas of constant motion from being picked up) to create a background and variance frame and then using those two frames and a difference frame (current frame – background frame), i get a resultant motion frame, the upper right frame in the below image.

What i need to do at this stage is “simply” extract the resultant silhouette, convert it into a numeric format: i intend on doing this by finding the centroid of the silhouette and get the lengths of a lines from the centroid to the edge in intervals of 2 degrees. Then I’ll take the 180 lengths I have and normalize them to remove any difference in scaling and then run them through a back-propagation neural network (BPN), first to train it and later for the actual classification.

Now the problem is the extraction of the the silhouettes. If i assume only a single object will be present its easy but as you can seen even a single object isn’t guaranteed to be represented by a single silhouette, due to the nature of current gray scale motion detection algorithms you tend to get a lot of breakage in the detected motion. So i have to extract and reconstruct these silhouettes. Sounds simple? Its not…

I’m running short on time and sitting and reading through stacks of journals for ready made algorithms which I’m still going to have to modify isn’t going to help, so with a bit of brain storming with my dad, we came up with a silly, simple solution. Anyone remember the classic computer science problem: traversing a maze? The solution? The right hand method, stick your right hand against the right wall and walk until you reach the end. I’m going to use this method to trace the silhouette outlines and then do a horizontal and vertical comparison on the separate silhouette outlines and then try a uninformed reconstruction. This is all just an idea right now and i’m not even sure if it will work and if it does whether it will be good enough to use in the final project.

I should actually get back to working on it. Hopefully by the end of the weekend, I’ll have more details and source code of my results. The worst part is that I’m not sure i have the processor time available for this technique, the current motion detection is expensive enough as it is. On my home X2 5400, it takes around 60-70% on the one core.