Logo recognition - how to improve performance

Question

I am working on a project of recognizing TV Channels. I am taking photos of the channels suck that i try to avoid the background and to take the sample from the center of th

Answer 1

It all depends on how you are calculating the channel probabilities from the templates. Are you using histogram of color or histogram of gradient, and then looking at the histogram difference between your templates and the test images?

Another approach would be to compute feature vectors from test images, such as the concatenation of a histogram of gradients and a histogram of color. Then, manually create a training database, in which you know the label (1, 2, 3, or 4 depending on which label is visible in the image) and you can feed the hard-coded labels, along with the histogram features, into a classifier routine. I recommend LIBSVM for this, and the scikits.learn implementation is easy to use for this in Python.

This will yield a support vector machine classifier that will compare the feature vector of new images with the support vectors from the training set, and determine the right label that is most likely present in the image. You can then fit a logistic model over top of this SVM if you want something that yields probabilities rather than just predicted labels.

Two good books to read to get started in this type of machine learning are Pattern Classification, by Duda, Hart, and Stork, and Pattern Recognition and Machine Learning by Bishop.

Some messy Python code that I wrote for implementing Poselets and Histogram of Oriented Gradients in Python can be found linked here; maybe you can grab some sections of code in there and it will be suitable for your task.