Ask for advices to character recognition
I have trained my character's classifier using Opencv's Support Vector Machine. But the outcome is not good as expected. My classifier always classify wrongly even for the training data.
The sample size I used in class 0-9 for each is about 15000, while the sample size of class A-Z for each is about 1500 - 4000. Now I want to try to use some of deep learning libraries, but I don't know whether it is practical to change. I have looked into Caffe and Eblearn, but they may not be suitable to integrate to my development platform Vs2005. And I like to develop in C++. Could anyone give some suggestions to me?
IMHO, your classifier is not good, your data is not good, if the prediction on the training data is bad. You should put just good samples in the training data; or use SVM::train_auto for well choosing the parameters.
Reading your question makes me wonder something: are you using any kind of feature extraction step prior to training/prediction with machine learning algorithms? Because you can even have 1.000.000.000 samples in your training set, but if you're trying to make classification based on pure pixels values, you're never going to have anything valid.
Thanks thdrksdfthmn and Lorena GdL. I get your idea right now. I just turn to use some feature descriptors. My current proposed method is to use some feature descriptors for grabbing the keypoints. According to some academic papers, they suggested that at each of the extracted keypoints, the HOG feature is calculated as an observed feature vector x in feature space, then turn them into BoW- based feature representation and finally use SVM for training. But I wonder if I'm doing it right. I post my code as follows.
I post my latest comment here since it is too long and I'm still a new user. My code@Lorena GdL , @thdrksdfthmn
http://answers.opencv.org/question/64... Please give a helping hand! @thdrksdfthmn