Intuition behind U-net vs FCN for semantic segmentation - neural-network

I don't quite understand the following:
In the proposed FCN for Semantic Segmentation by Shelhamer et al, they propose a pixel-to-pixel prediction to construct masks/exact locations of objects in an image.
In the slightly modified version of the FCN for biomedical image segmentation, the U-net, the main difference seems to be "a concatenation with the correspondingly cropped feature map from the contracting path."
Now, why does this feature make a difference particularly for biomedical segmentation? The main differences I can point out for biomedical images vs other data sets is that in biomedical images there are not as rich set of features defining an object as for common every day objects. Also the size of the data set is limited. But is this extra feature inspired by these two facts or some other reason?

FCN vs U-Net:
FCN
It upsamples only once. i.e. it has only one layer in the decoder
The original implementation github repo uses bilinear interpolation for upsampling the convoloved image. That is there is no learnable filter here
variants of FCN-[FCN 16s and FCN 8s] add the skip connections from lower layers to make the output robust to scale changes
U-Net
multiple upsampling layers
uses skip connections and concatenates instead of adding up
uses learnable weight filters instead of fixed interpolation technique

U-Net is built upon J. Long's FCN paper. A couple of differences is that the original FCN paper used the decoder half to upsample the classification (i.e the entire second half of the net is of depth C - number of classes)
U-Net's think of the second half as being in feature space and do the final classification at the end.
Nothing about it is special to bio-medical IMO

Related

How does upsampling in Fully Connected Convolutional network work?

I read several posts / articles and have some doubts on the mechanism of upsampling after the CNN downsampling.
I took the 1st answer from this question:
https://www.quora.com/How-do-fully-convolutional-networks-upsample-their-coarse-output
I understood that similar to normal convolution operation, the "upsampling" also uses kernels which need to be trained.
Question1: if the "spatial information" is already lost during the first stages of CNN, how can it be re-constructed in anyway ?
Question2: Why >"Upsampling from a small (coarse) featuremap deep in the network has good semantic information but bad resolution. Upsampling from a larger feature map closer to the input, will produce better detail but worse semantic information" ?
Question #1
Upsampling doesn't (and cannot) reconstruct any lost information. Its role is to bring back the resolution to the resolution of previous layer.
Theoretically, we can eliminate the down/up sampling layers altogether. However to reduce the number of computations, we can downsample the input before a layers and then upsample its output.
Therefore, the sole purpose of down/up sampling layers is to reduce computations in each layer, while keeping the dimension of input/output as before.
You might argue the down-sampling might cause information loss. That is always a possibility but remember the role of CNN is essentially extracting "useful" information from the input and reducing it into a smaller dimension.
Question #2
As we go from the input layer in CNN to the output layer, the dimension of data generally decreases while the semantic and extracted information hopefully increases.
Suppose we have the a CNN for image classification. In such CNN, the early layers usually extract the basic shapes and edges in the image. The next layers detect more complex concepts like corners, circles. You can imagine the very last layers might have nodes that detect very complex features (like presence of a person in the image).
So up-sampling from a large feature map close to the input produces better detail but has lower semantic information compared to the last layers. In retrospect, the last layers generally have lower dimension hence their resolution is worse compared to the early layers.

Face Recognition based on Deep Learning (Siamese Architecture)

I want to use pre-trained model for the face identification. I try to use Siamese architecture which requires a few number of images. Could you give me any trained model which I can change for the Siamese architecture? How can I change the network model which I can put two images to find their similarities (I do not want to create image based on the tutorial here)? I only want to use the system for real time application. Do you have any recommendations?
I suppose you can use this model, described in Xiang Wu, Ran He, Zhenan Sun, Tieniu Tan A Light CNN for Deep Face Representation with Noisy Labels (arXiv 2015) as a a strating point for your experiments.
As for the Siamese network, what you are trying to earn is a mapping from a face image into some high dimensional vector space, in which distances between points reflects (dis)similarity between faces.
To do so, you only need one network that gets a face as an input and produce a high-dim vector as an output.
However, to train this single network using the Siamese approach, you are going to duplicate it: creating two instances of the same net (you need to explicitly link the weights of the two copies). During training you are going to provide pairs of faces to the nets: one to each copy, then the single loss layer on top of the two copies can compare the high-dimensional vectors representing the two faces and compute a loss according to a "same/not same" label associated with this pair.
Hence, you only need the duplication for the training. In test time ('deploy') you are going to have a single net providing you with a semantically meaningful high dimensional representation of faces.
For a more advance Siamese architecture and loss see this thread.
On the other hand, you might want to consider the approach described in Oren Tadmor, Yonatan Wexler, Tal Rosenwein, Shai Shalev-Shwartz, Amnon Shashua Learning a Metric Embedding for Face Recognition using the Multibatch Method (arXiv 2016). This approach is more efficient and easy to implement than pair-wise losses over image pairs.

PCA on Sift desciptors and Fisher Vectors

I was reading this particular paper http://www.robots.ox.ac.uk/~vgg/publications/2011/Chatfield11/chatfield11.pdf and I find the Fisher Vector with GMM vocabulary approach very interesting and I would like to test it myself.
However, it is totally unclear (to me) how do they apply PCA dimensionality reduction on the data. I mean, do they calculate Feature Space and once it is calculated they perform PCA on it? Or do they just perform PCA on every image after SIFT is calculated and then they create feature space?
Is this supposed to be done for both training test sets? To me it's an 'obviously yes' answer, however it is not clear.
I was thinking of creating the feature space from training set and then run PCA on it. Then, I could use that PCA coefficient from training set to reduce each image's sift descriptor that is going to be encoded into Fisher Vector for later classification, whether it is a test or a train image.
EDIT 1;
Simplistic example:
[coef , reduced_feat_space]= pca(Feat_Space','NumComponents', 80);
and then (for both test and train images)
reduced_test_img = test_img * coef; (And then choose the first 80 dimensions of the reduced_test_img)
What do you think? Cheers
It looks to me like they do SIFT first and then do PCA. the article states in section 2.1 "The local descriptors are fixed in all experiments to be SIFT descriptors..."
also in the introduction section "the following three steps:(i) extraction
of local image features (e.g., SIFT descriptors), (ii) encoding of the local features in an image descriptor (e.g., a histogram of the quantized local features), and (iii) classification ... Recently several authors have focused on improving the second component" so it looks to me that the dimensionality reduction occurs after SIFT and the paper is simply talking about a few different methods of doing this, and the performance of each
I would also guess (as you did) that you would have to run it on both sets of images. Otherwise your would be using two different metrics to classify the images it really is like comparing apples to oranges. Comparing a reduced dimensional representation to the full one (even for the same exact image) will show some variation. In fact that is the whole premise of PCA, you are giving up some smaller features (usually) for computational efficiency. The real question with PCA or any dimensionality reduction algorithm is how much information can I give up and still reliably classify/segment different data sets
And as a last point, you would have to treat both images the same way, because your end goal is to use the Fisher Feature Vector for classification as either test or training. Now imagine you decided training images dont get PCA and test images do. Now I give you some image X, what would you do with it? How could you treat one set of images differently from another BEFORE you've classified them? Using the same technique on both sets means you'd process my image X then decide where to put it.
Anyway, I hope that helped and wasn't to rant-like. Good Luck :-)

Can neural network fail to learn a function? and How to choose better feature descriptors for pattern recognition?

I was working on webots which is an environment used to model, program and simulate mobile robots. Basically i have a small robot with a VGA camera, and it looks for simple blue coloured patterns on white walls of a small lego maze and moves accordingly
The method I used here was
​
Obtain images of the patterns from webots and save it in a location
in PC.
​​Detect the blue pattern, form a square enclosing the pattern
with atleast 2 edges of the pattern being part of the boundary of the
square.
​Resize it to 7x7 matrix(using nearest neighbour
interpolation algorithm)
The input to the network is nothing but the red pixel intensities of each of the 7x7 image(when i look at the blue pixel through a red filter it appears black so). The intensities of each pixel is extracted and the 7x7 matrix is then converted it to a 1D vector i.e 1x49 which is my input to the neural network. (I chose this characteristic as my input because it is 'relatively' less difficult to access this information using C and webots.​​)
I used MATLAB for this offline training method and I used a slower learning rate(0.06) to ensure parameter convergence and tested it on large and small datasets(1189 and 346 respectively). On all the numerous times I have tried, the network fails to classify the pattern.(it says the pattern belongs to all the 4 classes !!!! ) . There is nothing wrong with the program as I tested it out on the simpleclass_dataset in matlab and it works almost perfectly
Is it possible that the neural network fails to learn the function because of really poor data? (by poor data i mean that the datapoints corresponding to one sample of one class are very close to another sample belonging to a different class or something of that sort). Or can the neural network fail because of very poor feature descriptors?
Can anyone suggest a simpler method to extract features from the image(I am now shifting to MATLAB as I am now only concerned with simulations in webots and not the real robot). What sort of features can I choose? The patterns are very simple (L,an inverted L and its reflected versions are the 4 patterns)
Neural networks CAN fail to learn a function; this is most often caused by employing a network topology which is too simple to model the necessary function. A classic example of this case is attempting to learn an XOR function using a perceptron classifier, although it can even happen in multilayer neural nets sometimes; especially for complex tasks like image recognition. See my previous answer for a rough guide on how to select neural network parameters (ignore the convolution stuff if you want, although I would highly recommened looking into convolutional neural networks if you are still having problems).
It is a possiblity that there is too little seperability between classes, although I doubt that this is the case given your current features. Is there a reason that your network needs to allow an image to be four classifications simultaneously? If not, then perhaps you could classify the input as the output with the highest activation instead of all those with high activations.

Neural Network Input Bias in MATLAB

In Matlab (Neural Network Toolbox + Image Processing Toolbox), I have written a script to extract features from images and construct a "feature vector". My problem is that some features have more data than others. I don't want these features to have more significance than others with less data.
For example, I might have a feature vector made up of 9 elements:
hProjection = [12,45,19,10];
vProjection = [3,16,90,19];
area = 346;
featureVector = [hProjection, vProjection, area];
If I construct a Neural Network with featureVector as my input, the area only makes up 10% of the input data and is less significant.
I'm using a feed-forward back-propogation network with a tansig transfer function (pattern-recognition network).
How do I deal with this?
When you present your input data to the network, each column of your feature vector is fed to the input layer as an attribute by itself.
The only bias you have to worry about is the scale of each (ie: we usually normalize the features to the [0,1] range).
Also if you believe that the features are dependent/correlated, you might want to perform some kind of attribute selection technique. And in your case it depends one the meaning of the hProj/vProj features...
EDIT:
It just occurred to me that as an alternative to feature selection, you can use a dimensionality reduction technique (PCA/SVD, Factor Analysis, ICA, ...). For example, factor analysis can be used to extract a set of latent hidden variables upon which those hProj/vProj depends on. So instead of these 8 features, you can get 2 features such that the original 8 are a linear combination of the new two features (plus some error term). Refer to this page for a complete example