In the current Internet world, the numbers of digital images are growing exponentially. As a result, it is very tough to retrieve relevant objects for a given query point. For the past few years, researchers have been contributing different algorithms in the two most common machine learning categories to either cluster or classify images. There are several techniques of supervised classification images depending on the local or global feature representation of images, and on the metric used to calculate the distance (or similarity) between images. Recently many studies have shown the interest to learn a metric rather than use a simple metric a priori (e.g. Euclidean distance). This approach is described in the literature as metric learning. The main objective of this thesis is to use metric learning algorithm in the context of large-scale image classification. In this project, we use a metric learning algorithm which is driven by the nearest neighbors approach and has a competence to improve the generic k Nearest Neighbor (kNN) machine learning algorithm. Even though we get significant improvement on the performance of classification, the computation is very expensive due to the large dimensionality of our input dataset. Thus, we use the dimension reduction technique to reduce dimension and computation time as well. Nevertheless, due to the size of the database, classifying and searching a given query point using metric learning algorithm alone exhaustively is intractable.