How to create a single constant-length function vector from among variable image descriptors (SURF)

Question

My problem is this:

• I have 6 types of images or 6 classes. For example, cat, dog, bird, etc.
• For each type of image, I have many variations of this image. For example, a brown cat, a black dog, etc.
• I am currently using Vector Vector Machine (SVM) to classify images using one to another classification. I expand each image into one pixel vector and use it as a function vector for this image. I experience decent classification accuracy, but want to try something else.
• I want to use image descriptors, especially SURF functions, as a vector function for each image. This problem is that I can have only one feature vector for each image, and I am given a variable number of SURF functions from the function extraction process. For example, 1 cat picture can give me 40 SURF functions, while 1 dog picture will give me 68 SURF functions. I could select the n strongest functions, but I cannot guarantee that the selected SURF functions are those that describe my image (for example, it can focus on the background). There is also no guarantee that ANY SURF features are detected.

So my problem is how can I get a lot of observations (each of which is a SURF property vector) and “collapse” these observations into one feature vector, which describes the original image and can be transferred to SVM for training?

Thank you for your help!

Answer 1

Typically, SURF descriptors are quantized using a K-tool dictionary and aggregated into a single l1-normalized histogram. Thus, your entries in the SVM algorithm are now fixed in size.