I recently took up Machine Learning course on coursera and passed the course with decent scores. I used KNN,Logistic Regression,NN etc algorithms during the course, one assignment was to write learning algorithm for digits identification which I was able to complete. THe course ended with a case study of Photo OCR which really excited me, But I found it difficult to apply the basic algo taught in the course for this huge problem.So Can anyone suggest me some algorithms on Photo OCR?
I recently took up Machine Learning course on coursera and passed the course with
Share
The problem with image recognition is that it is highly sensitive to any change. They average human brain is able to extract certain features from the image will allow us to identify a given image even if certain image operations (such as skewing, rotating, etc) have been applied.
That being said, to my knowledge, Artifical Neural Netwoks are the most widely used (throwing in a hidden layer or two usually also helps). Another technique I have heard of is Wisard (or Wizard) but I cannot find anything about it. This technique basically breaks an image into sections and then you obtain a percentage of similarity when comparing the image segment with what you have in your knowledge base.
That being said, if I where you I would stick to neural networks plus a decent graphics manipulation library such as OpenCV (there are various wrappers for this, included Java and C#). The aim is the eliminate as much unneeded information as possible. In certain cases for instance, reducing the image to a gray scale or strictly black and white pixels helps.