24-03-2012, 04:26 PM
DIGITAL SIGNAL PROCESSING LAB
DIGITAL SIGNAL PROCESSING LAB report.pdf (Size: 298.28 KB / Downloads: 104)
INTRODUCTION:
Content-based image search is the application of computer vision techniques to the image retrieval problem. "Content-based" means that the search will analyze the actual contents of the image to retrieve images which contain the same data.
ALGORITHM :
A large dataset of images is collected that contain objects of various category for the training and testing purpose.
Training:
1. The SIFT (Scale Invariant Feature Transform) features of the images are found. These features are robust towards scaling, illumination changes, contrast changes and other affine transformations.
2. These features are than clustered together using K-Means clustering.
3. A vocabulary is then created using a Bag of Words model, where each cluster is considered as a word.
4. Different categories of images are chosen, and a classifier is trained for each category using the created vocabulary. The classifier becomes as strong as the number of images used to train it increases, in the sense that it better able to identify and discriminate objects.
5. The trained classifiers are stored as XML files for faster access.
RESULT:
The algorithm was implemented and tested in C. The results were satisfactory for some categories like car, bus, bicycle etc. For categories like cats, dogs, sheeps etc. some conflicting output was shown.
The accuracy of this method depends on the images that are used to train the classifiers. If the dataset is large enough and the images for each category are precise and accurate, that is without any other background objects, then the output is more satisfactory.