I. Simple CNN
II. Using pre-trained models
I. Simple CNN
II. Using pre-trained models
I remember back to the day when I started my PhD on iris recognition, there was only one iris recognition open source code from Libor Masek. His source code, written in Matlab, has been the baseline for generations of iris recognition coders. Recently there are a number of new open source codes come up. They are more mature and achieve close to state-of-the-art accuracy. I summarise them in a list here for your reference.
|Year/ Language||Approach||Performance (EER)|
|ICE 2005||MBGC portal||CASIA|
|Libor Masek [project]||2003
|Hough Circle + 1D Log-Gabor|
|OSIRIS 4.1 [project]||2013
|Least square Circle + 2D Gabor||1.09%|
|VASIR 2.2 [project]||2013
|Circle + 2D Gabor||3.5%||13.9% best quality frame
30.6% all frames
|Ellipse + 2D Gabor|
1D Log Gabor/
|Iris segmentation only|
|ICB 2016||IrisSeg: A Fast and Robust Iris Segmentation Framework for Non-Ideal Iris Images|
|Total-variation||ICCV 2015||An Accurate Iris Segmentation Framework under Relaxed Imaging Constraints using Total Variation Model|
|Geodesic Active Contours and GrabCut||PSIVT 2015||Iris Segmentation using Geodesic Active Contours and GrabCut|
Iris datasets to consider:
|CASIA-Iris-Distance||At a distance|
|ND-CrossSensor-Iris-2013||676||29,986 from LG4000 and 116,564 from LG2200||NIR||Cross Sensor|
|ND-TimeLapseIris-2012||23||6797||Time lapse 2004 to 2008|
|ND-Iris-Template-Aging-2008-2010||11,776||Time lapse 2008 to 2010|
|MBGC||NIR & NIR videos||One the move
|UBIRIS||261||11,102||Visible||On the move
At a distance
There is a website selling a number of iris recognition source codes based on different methods including Neural Networks, DCT, LBP, DFT and Genetics. All codes are on Matlab. I have not bought any to try out, but from their demonstration, it looks like the codes are just for demonstration, i.e. the accuracy is not the priority. Check it out to have a clear idea: http://www.irisrecognition.it/iris.asp or http://www.advancedsourcecode.com/iris.asp.
I know there are other source codes available out there. Let me know if you know any ones that are of interest to the iris recognition community.
II. UNDERSTAND convolutional neural networks
III. TRAIN image classification CNN models
Setup Caffe for use with Matlab
Install Caffe here
To use Caffe in Matlab, other than the installation steps in the above link, you need to compile Caffe for Matlab purpose:
Classifying images in Matlab using Caffe functions
Caffe is shipped with an Matlab example for classifying images in the /caffe/matlab/demo folder.This file is called: classification_demo.m.
Run Matlab. Sometime there is a error in linking the libstdc library, hence run the following from terminal before running Matlab:
$ export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libstdc++.so.6
The example can be run by two commands in Matlab:
im = imread(‘../../examples/images/cat.jpg’);
scores = classification_demo(im, 0); % Use 1 instead of 0 if you want to run on GPU
The code return scores for 1000 classes in the ImageNet dataset. You can run and debug the code to have a better insight into it.
Convolutional Neural Network (CNN) is a special type of feed-forward neural network dealing with images. One fundamental advantage of CNNs is the use of shared weight in convolutional layers, which means that the same filter (weights bank) is used for each pixel in the layer; this both reduces memory footprint and improves performance.
In this post, I will show how to test, understand and train this type of network.
I. CLASSIFYING images using pre-trained models
Caffe has already been packed with a number of examples, including one classification ipython notebook using CNN. This is a great example to start with. From terminal, navigate to the caffe installed path (usually in ~/caffe), run ipython notebook by typing:
This command calls ipython server to run and open up the web browser for working with ipython files. In the web browser just opened up, navigate to /caffe/examples/ and open file 00-classification.ipynb. Note that a ipython file is divided into multiple blocks so that you can run through each to debug the code. Use Shift+Enter to run through current block and move to the next block when the current finishes. Other shortcut keys can be found here.