Automating Image Registration for mulitple images (>2)

My Question is for a single image A, how to find its best matching image among thousands of image. Or to say, I concern the patterns appear in image A and where can I find it among all images I have (under a rigid, no distortion transformation)?

Hello, I now have, say, 10 point-set images (very simple binary images will around hundreds points) and what i want to do is to take first image as the reference image. And try to match it to the remaining 9 images.

I tried to do that with "Registration Estimator" in the image processing toolbox. However, I have plenty of images and I want to have an automatically matching. Can anyone tell me how to achieve this using Matlab?

Of course there will 9 matching but I want to find the best one. "Best" is the sense that I require a good quality score (can anyone tell me how Matlab calculate this? a target regression error?) and I dont mind to spend more time using maybe 2-3 algorithms like SURF, FAST to obtain a comprehensive search.

I also require the transformation will have no local distortion, if it does, I have a formula to describe the distortion. How can I give this restriction to Matlab?

One last question, can I regard the 1-quality score as my loss and treat it a machine learning problem because I have images >>9 (around multi thousands) and I dont want to search all images. I have to speed it up.

Please tell me if anywhere is not clear.

Thank you in advance and hope you have a good one.

Answer 1

[Aditya Mittal]

Hi,

I have done a similar project in past. And I think here you can use Bag of Visual words technique.

As per my understanding, you are having a query image, and you want to find the images related to that query image in your database. Using the technique of Bag of visual words, you will find the features from all the images in the database. And further create the clusters of the features.

For the process of feature extraction, I tried with different feature extraction techniques, out of which SIFT worked best for me. You can try with SURF as well.

Now you can read the query image and extract the features. This way the problem is now converted into a simple supervised machine learning question. Where you have classes (clusters of features) and data points (features of query image). Use any supervised learning algorithm to classify the model.

For finding the best relevant image, you can use R-precision, precision at K.

I hope this helps.

[Paul]

sorry for late reply. I am now handling this in other ways.

it didnt help, but I really appreciate your kindness. Hope you have a good day.