I have been working on an application that involves font recognition based on a users free hand drawing characters in Android Canvas.
In this application the user is asked to enter some predefined characters in a predefined order (A,a,B,c). Based on this, is there any way to show the very similar font which matches the user's hand writing.
I have researched on this topic found some papers & articles but most of them are recognizing font from a captured image. In that case they are having a lot of problems by segmenting paragraphs, individual letters and so on. But in my scenario I know what letter the user is drawing.
I have some knowledge in OpenCV and Machine Learning. Need help on how to proceed with this problem.
It is not exactly clear to me what you want to accomplish with your application but I assume that you are trying to output a font from a database of fonts that matches a users handwriting the most.
In Machine Learning this would be a classification problem. The number of classes will by equal to the number of different fonts in your database.
You could solve this with the help of a Convolutional neural network which are widely used for image and video recognition related tasks. If you've never implemented a CNN before I would suggest that you look up this resources to learn about Torch which is a easy-to-start-with toolkit to implement CNN's. (Of course there are more Frameworks such as: Tensor Flow, Caffe, Lasagne, ...)
Torch Homepage
Deep learning with Torch: 60 minutes blitz
Torch Cheatsheet
The main obstacle you will face is that Neural Networks need thousands of images (>100.000) to properly train them and to achieve satisfying results. Furthermore you do not only need the images but also a correct label for each image. Will say, you would need a training image such as a handwritten character and the corresponding font it matches the most out of your database as its label.
I would suggest that you read about so called transfer learning which can give you an initial boost as you do not need to set up a CNN model completely by yourself. In addition people have pre-trained such a model for a related task so that you safe extra time as you would not need to train it for many hours on a GPU. (see CUDA)
A great resource to start with is the paper: How transferable are features in deep neural networks?, which could be helpful for the stated reasons.
To get tons of training and testing data you can look up the following open datasets that provide all types of characters that can be helpful for your task:
Artificial Characters Data Set
UJI Pen Characters Data Set
The Chars74K dataset
Hand written - Datasets
A New Benchmark Dataset for Handwritten Character Recognition
For access to a lot of fonts and maybe even the possibility to create further datasets on your own you can have a look at Google Fonts.
You might find this article very interesting : https://erikbern.com/2016/01/21/analyzing-50k-fonts-using-deep-neural-networks/
Seems like a pretty straightforward deep learning supervised learning problem.
Generate a ton of randomly deformed samples for letters of each target font type, and train a convnet on that set?
The ideal would be to have a huge set of labeled, handwriting to font data, but that feels unlikely.
You could also use the generated, progressive to font code to take a bunch of handwritten samples, and transform them to look more like the font of your choice, as a dataset.
This is good place to start : https://github.com/fchollet/keras/blob/master/examples/mnist_cnn.py
Digit letter recognition with convnets.
This is quite a bit of work though if you haven't worked with that stuff before.
I would suggest using OCR library tesseract. Very well developed and mature. It also has support for training with other languages which you can use to train over a set of font.
Approach
Training:-
Take all 26(per alphabet) images for n fonts. Train tessaract over 26 A's, then 26 B's and soon.
Testing:-
Take a sentence and separate all characters.
For each character, find certainty score(supported in library) from Tesseract. Note, for character 'a, use the trained model on all 'a''s from different fonts.
For all characters, find best font using some metric (average, median, etc). For example: You can sum certainty score each font received for all characters and use the font which got max result.
Related
After trying several available eng.traineddata files with an Android app that employs Tesseract, I have less than stellar accuracy. Since my application will be using just a few fonts (font sizes, bold and regular), I thought I could get much better accuracy by building my own data. An example of the kind of thing (an 8.5x11 inch paper) that users will taking a picture of is here:
I have looked at jTessBoxEditor, but wondered if that was an appropriate path to investigate. And if so, I was unsure how to proceed with respect to a starting point, or to try from scratch. The font (which looks like Times New Roman) is very common, and didn't want to re-invent the wheel. I also wondered about how to treat the font on the two different color backgrounds.
Also, I wondered if I could just print-out ABC... abc... 123... in Times New Roman font and get that into a custom eng.traineddata file. If I understand correctly, you want the 'cleanest' data (i.e. no 'bad examples' of letters) in the source material used to train your system. But it would seem as if there would be a tutorial or procedure defined for how to build trained data for a specific font. If there is, it's been eluding me.
I would consider using machine learning, but so you don't have to do it on your own, look at Tensorflow Mobile. This is a version that is for mobile devices, and to help with character recognition you can look at this article.
To train any neural net a set of training data along with correct
outputs must be provided. In this case this will be a set of 128x64
images along with the expected output.
This will help you easily implement a solution to recognize the characters, and by going with this approach you can extend to more fonts if you desire by just doing more training.
Right now I am working on an android and ios project, and this app will allow people to visit a place(like a park), and it will contain a feature that can let user to know what kind of flower by scanning the flower with their phone, and thw app will match the flower in the database.
Well, the problem I have is, there are ways I can do this template matching thing, and now I have two ways:
1.use OpenCV for both android and iosSimple RGB analysis using java
2.Simple RGB analysis using java
But both of these ways seems not so peefect. At first, I searched for template matching, and there seems to be a lot of people talking about OpenCV. But I realized that OpenCV is only supported up to android 4.0, so I tried to find a way that makes this more simple, which is using a simple JAVA code with RGB matching. But here's another problem: simple RGB matching is not a great idea for flowers, and even though every flowers has different colors, different cameras may cause the RGB result to be very different.
Is there any better way to match flowers?
(Sorry for my bad english)
I would like to make an Android Application that captures an image and searches it for coins and paper notes and then determines the value of the money in the image.
Additionally, the output of the system will be such that it can be understood by a blind person.
What functions and techniques in openCV would suit these tasks?
What limitations and development hurdles can I expect?
Assuming you already know how to program android apps,you need to do the following:-
Download the OpenCV SDK and set it up with the IDE.
Recognising shapes will be a huge part of your project, see the contour detection example that is present with the SDK examples. Your primary goal will be detecting a circle. Later you need to adapt your algorithm according to the currency. This will be of particular interest to you.
Learn the different image processing techniques like thresholding for better accurate results. Understanding what a Mat object is and how it can be manipulated is important.
Finally improve the accuracy of your algorithm, as sometimes lighting conditions make the difference between a good review and a dissatisfied user.
I have successfully integrated tesseract into my android app and it reads whatever the image that I capture but with very less accuracy. But most of the time I do not get the correct text after capturing because some text around the region of interest is also getting captured.
All I want to read is all text from a rectangular area, accurately, without capturing the edges of the rectangle. I have done some research and posted on stackoverflow about this two times, but still did not get a happy result!
Following are the 2 posts that I made:
https://stackoverflow.com/questions/16663504/extract-text-from-a-captured-image?noredirect=1#comment23973954_16663504
Extracting information from captured image in android
I am not sure whether to go ahead with tesseract or use openCV
Including the many links and answers from others, I think it's good to take a step back and note that there are actually two fundamental steps to optical character recognition (OCR):
Text Detection: This is the title and focus of your question, and it is concerned with localizing regions in an image that contain text.
Text Recognition: This is where the actual recognition happens, where the localized image regions from detection get segmented character-by-character and classified. This is also where tools like Tesseract come into play.
Now, there are also two general settings in which OCR is applied:
Controlled: These are images taken from a scanner or similar in-nature where the target is a document and things like perspective, scale, font, orientation, background consistency, etc are pretty docile.
Uncontrolled/Scene: These are the more natural and in-the-wild photos, e.g. those taken from a camera, where you are trying to recognize a street sign, shop name, etc.
Tesseract as-is is most applicable to the "controlled" setting. And in general, but for scene OCR especially, "re-training" Tesseract will not directly improve detection, but may improve recognition.
If you are looking to improve scene text detection, see this work; and if you are looking at improving scene text recognition, see this work. Since you asked about detection, the detection reference uses maximally stable extremal regions (MSER), which has a plethora of implementation resources, e.g. see here.
There's also a text detection project here specifically for Android too:
https://github.com/dreamdragon/text-detection
As many have noted, keep in mind that recognition is still an open research challenge.
The solution to improving the OCR output is to
either use more training data to train it better
filter it's input using some Linear Filter (grayscaling, high-contrasting, blurring)
In the chat we posted a number of links describing filtering techniques used in OCRing, but sample code wasn't posted.
Some of the links posted were
Improving input for OCR
How to train Tesseract
Text enhancement using asymmetric filters <-- this paper is easily found on google, and should be read fully as it quite clearly illustrates and demonstrates necessary steps before OCR-processing the image.
OCR Classification
We're currently working on an android ocr app using opencv.pre-processing ,segmentation ,Feature extraction steps are done. Classification is the remaining step and we're stuck ..We're using a DB table which is filled with each letter features ..Firstly we had only 1 feature per letter and we used euclidean distance ,but results wasn't accurate and more features needed to be obtained and so we did.The problem now is we have 7 features per letter and absolutely no idea of how to classify i/p based on them..some have recommended using knn ,but we can't figure out how and the opencv documentation in that part ain't clear ..so if anybody can help it wud be great.
Thanks in advance
Briefly and without discussing the details. Vector space comes in handy here. You need to build a feature vector
<feature1, feature2, feature3.. featureN> for each of the instances in your training set.
From each of these images you extract features that you think or you read in the research articles are important for image classification. For example you can do centroid, Gaussian blur, histograms, etc.
Once you have these values linear algebra comes into play with some classification algorithm: knn, svm, naive bayes etc that you run on your training set, that is you build your model.
If the model is ready you run it on your test set.
Use cross validation for more comprehensive results.
For more details check the course notes:
http://www.inf.ed.ac.uk/teaching/courses/iaml/slides/knn-2x2.pdf
or
http://www.inf.ed.ac.uk/teaching/courses/inf2b/lectureSchedule.html
would like to add that OpenCV may not have the sort of classifiers you might prefer.
There are several libraries out there, though you may have to see which works best when on a mobile platform. Could you give some details on the features you are using?
The simplest KNN (k-nearest neighbors) measure would be to find the Euclidean distance in n dimensions (for an n-dimensional feature vector) between the input sample's features and each of the vectors in your DB table. Also explore Mahalanobis distance (used to measure distance between a point and a dataset/class) if you have multiple classes and the input image is to be classified as one such 'type' or 'class' of image.
As #matcheek mentioned, more sophistication can be possible using machine learning techniques such as SVM, Neural Nets, etc. However first you might consider a simpler thing like kNN, considering its a mobile platform which may limit the computational complexity.