So I have an implementation for a neural network that I followed on Youtube. The guy uses SGD (Momentum) as an optimization algorithm and hyperbolic tangent as an activation function. I already changed the transfer function to Leaky ReLU (for the hidden layers) and Sigmoid (for the output layer).
But now I decided I should also change the optimization algorithm to Adam. And I ended up searching for SGD (Momentum) on Wikipedia for a deeper understanding of how it works and I noticed something's off. The formula the guy uses in the clip is different from the one on Wikipedia. And I'm not sure if that's a mistake, or not... The clip is one hour long, but I'm not asking you to watch the entire video, however I'm intrigued by the 54m37s mark and the Wikipedia formula, right here:
https://youtu.be/KkwX7FkLfug?t=54m37s
https://en.wikipedia.org/wiki/Stochastic_gradient_descent#Momentum
So if you take a look at the guy's implementation and then at the Wikipedia link for SGD (Momentum) formula, basically the only difference is in delta weight's calculation.
Wikipedia states that you subtract from the momentum multiplied by the old delta weight, the learning rate multiplied by the gradient and the output value of the neuron. Whereas in the tutorial, instead of subtracting the guy adds those together. However, the formula for the new weight is correct. It simply adds the delta weight to the old weight.
So my question is, did the guy in the tutorial make a mistake, or is there something I am missing? Because somehow, I trained a neural network and it behaves accordingly, so I can't really tell what the problem is here. Thanks in advance.
I have seen momentum implemented in different ways. Personally, I followed this guide in the end: http://ruder.io/optimizing-gradient-descent
There, momentum and weights are updated separately, which I think makes it clearer.
I do not know enought about the variables in the video, so I am not sure about that, but the wikipedia version is deffinetly correct.
In the video, the gradient*learning_rate gets added instead of subtracted, which is fine if you calculate and propagate your error accordingly.
Also, where in the video says "neuron_getOutputVal()*m_gradient", if it is as I think it is, that whole thing is considered the gradient. What I mean is that you have to multiplicate what you propagate times the outputs of your neurons to get the actual gradient.
For gradient descent without momentum, once you have your actual gradient, you multiply it with a learning rate and subtract (or add, depending on how you calculated and propagated the error, but usually subtract) it from your weights.
With momentum, you do it as it says in the wikipedia, using the last "change to your weights" or "delta weights" as part of your formula.
Related
I'm trying to implement the LineTrace algorithm described in this article:
Linetrace Generative Art
Particularly where it says:
To trace the outline you can sample some of the nearby edges on the previous line, calculate the average direction of those edges and add a vertex to the current line along that direction. Then add some random motion to mimic free hand drawing. This seems to work quite well for a while, but there is some "inertia" that can be seen in the results—the shape adapts too slowly.
The amount of noise you add to each vertex is crucial. This noise is what drives the whole system to make interesting shapes since the tracing behaviour is always forced to attempt to replicate both the general movement and some of the random jitter as it progresses.
I'm trying to do this in processing, and since I'm new to processing and hazy on how vectors, edges and directions work, I don't have any idea how to start to code. I would be greatly appreciative for some sample code, anything to help me get underway. I'm also curious by what he means by "add some random motion to mimic free hand drawing", is he incorporating perlin noise somehow? Thanks in advance.
I'm looking for some algorithm (preferably if source code available)
for image registration.
Image deformation can't been described by homography matrix(because I think that distortion not symmetrical and not
homogeneous),more specifically deformations are like barrel/distortion and trapezoid distortion, maybe some rotation of image.
I want to obtain pairs of pixel of two images and so i can obtain representation of "deformation field".
I google a lot and find out that there are some algorithm base on some phisics ideas, but it seems that they can converge
to local maxima, but not global.
I can affort program to be semi-automatic, it means some simple user interation.
maybe some algorithms like SIFT will be appropriate?
but I think it can't provide "deformation field" with regular sufficient density.
if it important there is no scale changes.
example of complicated field
http://www.math.ucla.edu/~yanovsky/Research/ImageRegistration/2DMRI/2DMRI_lambda400_grid_only1.png
What you are looking for is "optical flow". Searching for these terms will yield you numerous results.
In OpenCV, there is a function called calcOpticalFlowFarneback() (in the video module) that does what you want.
The C API does still have an implementation of the classic paper by Horn & Schunck (1981) called "Determining optical flow".
You can also have a look at this work I've done, along with some code (but be careful, there are still some mysterious bugs in the opencl memory code. I will release a corrected version later this year.): http://lts2www.epfl.ch/people/dangelo/opticalflow
Besides OpenCV's optical flow (and mine ;-), you can have a look at ITK on itk.org for complete image registration chains (mostly aimed at medical imaging).
There's also a lot of optical flow code (matlab, C/C++...) that can be found thanks to google, for example cs.brown.edu/~dqsun/research/software.html, gpu4vision, etc
-- EDIT : about optical flow --
Optical flow is divided in two families of algorithms : the dense ones, and the others.
Dense algorithms give one motion vector per pixel, non-dense ones one vector per tracked feature.
Examples of the dense family include Horn-Schunck and Farneback (to stay with opencv), and more generally any algorithm that will minimize some cost function over the whole images (the various TV-L1 flows, etc).
An example for the non-dense family is the KLT, which is called Lucas-Kanade in opencv.
In the dense family, since the motion for each pixel is almost free, it can deal with scale changes. Keep in mind however that these algorithms can fail in the case of large motions / scales changes because they usually rely on linearizations (Taylor expansions of the motion and image changes). Furthermore, in the variational approach, each pixel contributes to the overall result. Hence, parts that are invisible in one image are likely to deviate the algorithm from the actual solution.
Anyway, techniques such as coarse-to-fine implementations are employed to bypass these limits, and these problems have usually only a small impact. Brutal illumination changes, or large occluded / unoccluded areas can also be explicitly dealt with by some algorithms, see for example this paper that computes a sparse image of "innovation" alongside the optical flow field.
i found some software medical specific, but it's complicate and it's not work with simple image formats, but seems that it do that I need.
http://www.csd.uoc.gr/~komod/FastPD/index.html
Drop - Deformable Registration using Discrete Optimization
Found this very interesting code on total variation filter tvmfilter
The additional functions this code uses are very confusing but the denoising is far better than all the filters i have tried so far
i have figured out the code on my own :)
His additional function "tv" denoises with the ROF model which has been a major research topic for two decades now. See http://www.ipol.im/pub/algo/g_tv_denoising/ for a summary of current methods.
Briefly, the idea behind ROF is to approximate the given noisy image with a piecewise constant image by solving an optimization which penalizes the total variation (ie l1-norm of the gradient) of the image.
The reason this performs well is that the other denoising methods you are probably working with denoise by smoothing the image via convolution with a Gaussian (ie penalizing the l2-norm of the gradient (ie solving the heat equation on the image) ). While fast to compute, denoising by smoothing blurs edges and thus results in poor image quality. l1-norm optimization preserves edges.
It's not clear how Guy solves the tv problem in that code you linked. He references the original ROF paper so it's possible that he's just using the original method (gradient descent) which is quite slow to converge. I suggest you give this code/paper a try: http://www.stanford.edu/~tagoldst/Tom_Goldstein/Split_Bregman.html as it's probably faster than the .m file you are using.
Also, as was mentioned in the comments, you will get better denoising (ie higher SNR) using nonlocal means. However, it will take much longer for the nonlocal means algorithm to work as it requires that you search the entire image for similar patches and compute weights based on them.
I am thinking of implement a image processing based solution for industrial problem.
The image is consists of a Red rectangle. Inside that I will see a matrix of circles. The requirement is to count the number of circles under following constraints. (Real application : Count the number of bottles in a bottle casing. Any missing bottles???)
The time taken for the operation should be very low.
I need to detect the red rectangle as well. My objective is to count the
items in package and there are no
mechanism (sensors) to trigger the
camera. So camera will need to capture
the photos continuously but the
program should have a way to discard
the unnecessary images.
Processing should be realtime.
There may be a "noise" in image capturing. You may see ovals instead of circles.
My questions are as follows,
What is the best edge detection algorithm that matches with the given
scenario?
Are there any other mechanisms that I can use other than the edge
detection?
Is there a big impact between the language I use and the performance of
the system?
AHH - YOU HAVE NOW TOLD US THE BOTTLES ARE IN FIXED LOCATIONS!
IT IS AN INCREDIBLY EASIER PROBLEM.
All you have to do is look at each of the 12 spots and see if there is a black area there or not. Nothing could be easier.
You do not have to do any edge or shape detection AT ALL.
It's that easy.
You then pointed out that the box might be rotatated, things could be jiggled. That the box might be rotated a little (or even a lot, 0 to 360 each time) is very easily dealt with. The fact that the bottles are in "slots" (even if jiggled) massively changes the nature of the problem. You're main problem (which is easy) is waiting until each new red square (crate) is centered under the camera. I just realised you meant "matrix" literally and specifically in the sentence in your original questions. That changes everything totally, compared to finding a disordered jumble of circles. Finding whether or not a blob is "on" at one of 12 points, is a wildly different problem to "identifying circles in an image". Perhaps you could post an image to wrap up the question.
Finally I believe Kenny below has identified the best solution: blob analysis.
"Count the number of bottles in a bottle casing"...
Do the individual bottles sit in "slots"? ie, there are 4x3 = 12 holes, one for each bottle.
In other words, you "only" have to determine if there is, or is not, a bottle in each of the 12 holes.
Is that correct?
If so, your problem is incredibly easier than the more general problem of a pile of bottles "anywhere".
Quite simply, where do we see the bottles from? The top, sides, bottom, or? Do we always see the tops/bottoms, or are they mixed (ie, packed top-to-tail). These issues make huge, huge differences.
Surf/Sift = overkill in this case you certainly don't need it.
If you want real time speed (about 20fps+ on a 800x600 image) I recommend using Cuda to implement edge detection using a standard filter scheme like sobel, then implement binarization + image closure to make sure the edges of circles are not segmented apart.
The hardest part will be fitting circles. This is assuming you already got to the step where you have taken edges and made sure they are connected using image closure (morphology.) At this point I would proceed as follows:
run blob analysis/connected components to segment out circles that do not touch. If circles can touch the next step will be trickier
for each connected componet/blob fit a circle or rectangle using RANSAC which can run in realtime (as opposed to Hough Transform which I believe is very hard to run in real time.)
Step 2 will be much harder if you can not segment the connected components that form circles seperately, so some additional thought should be invested on how to guarantee that condition.
Good luck.
Edit
Having thought about it some more, I feel like RANSAC is ideal for the case where the circle connected components do touch. RANSAC should hypothetically fit the circle to only a part of the connected component (due to its ability to perform well in the case of mostly outlier points.) This means that you could add an extra check to see if the fitted circle encompasses the entire connected component and if it does not then rerun RANSAC on the portion of the connected component that was left out. Rinse and repeat as many times as necessary.
Also I realize that I say circle but you could just as easily fit an ellipse instead of circles using RANSAC.
Also, I'd like to comment that when I say CUDA is a good choice I mean CUDA is a good choice to implement the sobel filter + binirization + image closing on. Connected components and RANSAC are probably best left to the CPU, but you can try pushing them onto CUDA though I don't know how much of an advantage a GPU will give you for those 2 over a CPU.
For the circles, try the Hough transform.
other mechanisms: dunno
Compiled languages will possibly be faster.
SIFT should have a very good response to circular objects - it is patented, though. GLOHis a similar algorithm, but I do not know if there are any implementations readily available.
Actually, doing some more research, SURF is an improved version of SIFT with quite a few implementations available, check out the links on the wikipedia page.
Sum of colors + convex hull to detect boundary. You need, mostly, 4 corners of a rectangle, and not it's sides?
No motion, no second camera, a little choice - lot of math methods against a little input (color histograms, color distribution matrix). Dunno.
Java == high memory consumption, Lisp == high brain consumption, C++ == memory/cpu/speed/brain use optimum.
If the contrast is good, blob analysis is the algorithm for the job.
I want to implement the two above mentioned image resampling algorithms (bicubic and Lanczos) in C++. I know that there are dozens of existing implementations out there, but I still want to make my own. I want to make it partly because I want to understand how they work, and partly because I want to give them some capabilities not found in mainstream implementations (like configurable multi-CPU support and progress reporting).
I tried reading Wikipedia, but the stuff is a bit too dry for me. Perhaps there are some nicer explanations of these algorithms? I couldn't find anything either on SO or Google.
Added: Seems like nobody can give me a good link about these topics. Can anyone at least try to explain them here?
The basic operation principle of both algorithms is pretty simple. They're both convolution filters. A convolution filter that for each output value moves the convolution functions point of origin to be centered on the output and then multiplies all the values in the input with the value of the convolution function at that location and adds them together.
One property of convolution is that the integral of the output is the product of the integrals of the two input functions. If you consider the input and output images, then the integral means average brightness and if you want the brightness to remain the same the integral of the convolution function needs to add up to one.
One way how to understand them is to think of the convolution function as something that shows how much input pixels influence the output pixel depending on their distance.
Convolution functions are usually defined so that they are zero when the distance is larger than some value so that you don't have to consider every input value for every output value.
For lanczos interpolation the convolution function is based on the sinc(x) = sin(x*pi)/x function, but only the first few lobes are taken. Usually 3:
lanczos(x) = {
0 if abs(x) > 3,
1 if x == 0,
else sin(x*pi)/x
}
This function is called the filter kernel.
To resample with lanczos imagine you overlay the output and input over eachother, with points signifying where the pixel locations are. For each output pixel location you take a box +- 3 output pixels from that point. For every input pixel that lies in that box, calculate the value of the lanczos function at that location with the distance from the output location in output pixel coordinates as the parameter. You then need to normalize the calculated values by scaling them so that they add up to 1. After that multiply each input pixel value with the corresponding scaling value and add the results together to get the value of the output pixel.
Because lanzos function has the separability property and, if you are resizing, the grid is regular, you can optimize this by doing the convolution horizontally and vertically separately and precalculate the vertical filters for each row and horizontal filters for each column.
Bicubic convolution is basically the same, with a different filter kernel function.
To get more detail, there's a pretty good and thorough explanation in the book Digital Image Processing, section 16.3.
Also, image_operations.cc and convolver.cc in skia have a pretty well commented implementation of lanczos interpolation.
While what Ants Aasma says roughly describes the difference, I don't think it is particularly informative as to why you might do such a thing.
As far as links go, you are asking a very basic question in image processing, and any decent introductory textbook on the subject will describe this. If I remember correctly, Gonzales and Woods is decent on it, but I'm away from my books and can't check.
Now on to the particulars, it should help to think about what you are doing fundamentally. You have a square lattice of measurements that you want to interpolate new values for. In the simple case of upsampling, lets imagine you want a new measurement in between every one that you already have (e.g. double the resolution).
Now you won't get the "correct" value, because in general you don't have that information. So you have to estimate it. How to do this? A very simple way would be to linearly interpolate. Everyone knows how to do this with two points, you just draw a line between them, and read the new value off the line (in this case, at the half way point).
Now an image is two dimensional, so you really want to do this in both the left-right and up-down directions. Use the result for your estimate and voila you have "bilinear" interpolation.
The main problem with this is that it isn't very accurate, although it's better (and slower) than the "nearest neighbor" approach which is also very local and fast.
To address the first problem, you want something better than a linear fit of two points, you want to fit something to more data points (pixels), and something that can be nonlinear. A good trade off on accuracy and computational cost is something called a cubic spline. So this will give you a smooth fit line, and again you approximate your new "measurement" by the value it takes in the middle. Do this in both directions and you've got "bicubic" interpolation.
So that's more accurate, but still heavy. One way to address the speed issue is to use a convolution, which has the nice property that in the Fourier domain, it's just a multiplication, so we can implement it quite quickly. But you don't need to worry about the implementation to understand that the convolution result at any point is one function (your image) being integrated in product another, typically much smaller support (the part that is non-zero) function called the kernel), after that kernel has been centered over that particular point. In the discrete world, these are just sums of the products.
It turns out that you can design a convolution kernel that has properties quite like the cubic spline, and use that to get a fast "bicubic"
Lancsoz resampling is a similar thing, with slightly different properties in the kernel, which primarily means they will have different characteristic artifacts. You can look up the details of these kernel functions easily enough (I'm sure wikipedia has them, or any intro text). The implementations used in graphics programs tend to be highly optimized and sometimes have specialized assumptions which make them more efficient but less general.
I would like suggest the following article for a basic understanding of different image interpolation methods image interpolation via convolution. If you want to try more interpolation methods, the imageresampler is a nice open source project to begin with.
In my opinion image interpolation can be understood from two aspects, one is from function fitting perspective, and one is from convolution perspective. For example, the spline interpolation explained in image interpolation via convolution is well explained from function fitting perspective in Cubic interpolation.
Additionally, image interpolation is always related to a specific application, for example image zooming, image rotation and so on. In fact for a specific application, image interpolation can be implemented i.n a smart way. For example, image rotation can be implemented via a three-shearing method, and during each shearing operation different one-dimension interpolation algorithms can be implemented.