Making dataset for signature recognition?
3 visualizzazioni (ultimi 30 giorni)
Mostra commenti meno recenti
Mansoor ahmadi
il 25 Gen 2015
Commentato: Fadi Alsuhimat
il 9 Feb 2019
Hello!
I am working on signature recognition system using neural network, this system recognize 360 signature images from 30 person, for each person 12 signature (8 genuine and 4 forge). How can I make dataset for training and testing for neural network to recognize genuine and forge?
Can someone help me!
heeeeeeeeeeeeeeeelp!!!!!!!
thanks in advance.
0 Commenti
Risposta accettata
Image Analyst
il 26 Gen 2015
Modificato: Image Analyst
il 26 Gen 2015
I'd probably get 360 people to hand write their signatures 8 times on white paper and then scan them in with a flatbed scanner to make 360*8 image files all stored in a "genuine" folder. Then get them people to "forge" signatures of 4 other people and scan and store those in a "forged" folder. When you assign people to forge signatures, make sure that every person has their signature forged 4 times by 4 different other people.
You can train with half and test with half or whatever fraction you want. Or you can use "Leave one out" http://en.wikipedia.org/wiki/Neighbourhood_components_analysis#Leave-one-out_.28LOO.29_classification
To determine accuracy of your NN algorithm, you might construct an ROC curve http://en.wikipedia.org/wiki/Receiver_operating_characteristic
6 Commenti
Image Analyst
il 26 Gen 2015
- I would just encode the name of the person and whether it's genuine or forged into the filename, either the base filename (like the person's name), or the folder name (like whether it's genuine or forged). When you read the image you can parse the filename to find out who it is and whether it's genuine of forged. You can certainly use dir() to get the names of all the files and then parse the filenames and store the data in a table or structure array if you want.
- I have no idea. You'll need to wait for Greg Heath to answer the NN question.
Più risposte (1)
Greg Heath
il 27 Gen 2015
I assume you have a technique for extracting features out of the images. If not, you'll have to search the net, including comp.ai.neural-nets as well as the NEWSGROUP and ANSWERS.
The best classification technique I can think of:
A 31 class classifier with thresholds trained on all signatures with a target matrix based on class indices 1:31 converted to 31-dimensional {0,1} unit vectors via function ind2vec. The input is assigned to the class associated with the maximum output PROVIDED the output exceeds the class specific threshold.
Therefore, if max(y) = y(10) and y(10) >= Thresh(10), then assignedclass = vec2ind(y) otherwise there is no classification.
To be clear, all forgeries are associated with target [zeros(30,1); 1]
The class-dependent thresholds are chosen via trial and error.
Hope this helps
Thank you for formally accepting my answer
Greg
PS A 60 class classifier might be better but your data base doesn't look large enough. classifier might be based on 60 classes.
2 Commenti
Fadi Alsuhimat
il 9 Feb 2019
I have same problem now, can you help me if you get the answer?
with my regard
Vedere anche
Categorie
Scopri di più su Data Extraction in Help Center e File Exchange
Prodotti
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!