70 likes | 141 Views
Searching for images is crucial due to the exponential growth of image databases. Content-Based Image Retrieval (CBIR) organizes images without user tags. The bottleneck is the classification algorithm, particularly k-means, an unsupervised method for organizing features. Scaling is vital for handling vast databases like Flickr and Facebook, achieved through parallelizing the k-means algorithm. The general k-means algorithm separates vectors into k clusters. Parallelization minimizes the computational bottleneck, with comparisons between OpenMP and MPI implementations. Future work includes parameter auto-tuning, integration with the CBIR system at Berkeley PAR lab, and testing on larger datasets.
E N D
Scaling Content Based Image Retrieval Systems Christine Lo, Sushant Shankar, ArunVijayvergiya CS 267
Motivation • Finding an efficient way to search for images has been increasingly important, especially since image databases are growing at an unprecedented rate. For example, there are about 550,000 images uploaded to Facebook each second. • Content Based Image Retrieval (CBIR) offers a way to classify images based on implicit criteria rather than user generated tags. This will make large image databases more organized and more searchable. • The bottleneck of the CBIR system is the classification algorithm. k-means is one of the classification methods we use for CBIR. We chose k-means because it is an unsupervised learning technique that will allow us to organize and classify unlabeled features. • Because of the size of image databases such as Flickr and Facebook, it is important to scale the classification algorithm to handle a large number of features. We accomplish this by parallelizing the k-means algorithm.
K-means Algorithm • General k-means Algorithm • Takes as input a list of vectors and separates them into k clusters. • Parallelization • We parallelize the k-means algorithm to minimize the computational bottleneck of the CBIR system. We compare two implementations of this, an OpenMP and an MPI implementation.
Future Work • Auto-tune parameters for best results • Integrate clustering code with CBIR system at Berkeley PAR lab • Test on larger datasets such as Flickr and Facebook