Presentation on theme: "16-20 Dec 2014, IDRBT, Hyderabad By Biswajit Halder 1, Rajkumar Darbar 2, Utpal Garain 3 and Abhoy Ch. Mondal 1 1Dept. Of Computer Science, University."— Presentation transcript:
16-20 Dec 2014, IDRBT, Hyderabad By Biswajit Halder 1, Rajkumar Darbar 2, Utpal Garain 3 and Abhoy Ch. Mondal 1 1Dept. Of Computer Science, University Of Burdhwan, Burdhwan, W.B., India 2 School of Information Technology, IIT, Kharagpur, India 3 Indian Statistical Institute, Kolkata, India Analysis of fluorescent paper pulps for detecting Counterfeit Indian paper money 10 th International Conference of Information System and Security
Counterfeiting of bank notes Press Release -  “1 st half of 2012 - Biannual information on euro banknote counterfeiting - on average 14.6 billion” –by European central bank on 16 July 2012  “In the 2011-2012 Fiscal Year, the Reserve Bank of India detected 521,155 counterfeit banknotes in India. The number of detects was 20 percent higher than 2010-2011 Fiscal Year.” – by Hindustan Times, August 26, 2012  "During 2013-14, the detection of counterfeit notes of Rs 1,000 and Rs 100 increased by 11.8% and 9.8 percent respectively, whereas that of Rs 500 denomination decreased by 10.3%," the central bank said in its annual report
Automatic Authentication Required Automatic Authentication process manual authentication of security features time consuming process unattractive solution Present Drawback -Not consider paper based security -Paper quality check depend on physical contacts -Question raise on Speed and accuracy
Color optical pulp based paper security -Embedding certain special ingredients at the time of manufacturing -Security fibers may be metallic or photo- chromic. -luminescent Under ultraviolet (UV) ray
Motivation -The pattern recognisation and artificial intelligence principles. -Ideas of feature extraction done from rice grain detection
Compare with paper pulp detection Similarity – i) shape ii) color ii) distribution Dissimilarity – i) other artwork involve ii) factor of non pulp elements
overall approach It’s divided into a number of stages:- (i) Detect pulps in a UV scanned banknote under identification and verification phase (ii) Extract features from the detected pulps (iii) Train a NN classifier based training samples that include both genuine and fake notes. (iv) Trained classifier configured as 2-class (genuine vs. fake) problem.
Algorithm-I: PULP IDENTIFICATION Step 1: Acquire the currency note image (RGB) by UV light Step 2: Image Complement (RGB -> CMY) Step 3: Extract cyan image Step 4: Apply median filter Step 5: Convert binary image by OTSU thresholding Step 6: Connected component labelling of background pixels Step 7: Compute centroid of each component
Algorithm-II: ELIMINATION OF NON- PULP ELEMENTS Step 1: Around each centroid as detected at Step 7 of Al- gorithm-I, mxm sub-image is cropped from the initial RGB image Step 2: For each such sub-image, compute Gray Level Co- occurrence Matrix (GLCM)  under consideration of two adjacent pixels on four directions: 0, 45, 90, and 135 degree Step 3: Generate texture level four statistical features i.e. contrast, correlation, energy and homogeneity from each co-occurrence matrix  Step 4: Configure an artificial neural Network (ANN) for discriminating pulps from non-pulp particles
Detection (a) Detected pulps after execution of Algorithm-I (b) Pulps after elimination non-pulp elements by Algorithm-II.
Features – (i) Area (f1): The number of pixels inside pulp identified by a connected component. (ii) Rectangular aspect ratio (f2): The ratio of the length and width of the rectangle bounding the pulp. (iii) Pulp Aspect ratio (f3): The ratio of the lengths of the major and minor axes. (iv) Shape factor (f4): Root means squared deviation of diameter of the pulp. (v) Avg. HSI (f5, f6, f7) (vi) Variances of HIS (f8, f9, f10)
Training of the Classifier Multilayer back propagation neural network (BPNN) used under following specification – -Input, 10 node for 10 features and output, 1 node -No. of node in hidden layer computed by where I and O, input and output node Y is no of pattern in training set - gradient descent method used for optimization
Pulp level authentication using Neural Network.
Document level authentication using Neural Network. It No.Genuine SampleFake Sample GFCFGC 147214532 249104613 344334730 448024325 Avg.94%3% 90.50%4.50%5%
Behavior of the neural net in classifying pulps (a) confusion matrix (b) ROC plot