Image Matching Using Dimensionally Reduced Embedded Earth Mover’s Distance

Abstract

Finding similar images to a given query image can be computed by different distance measures. One of the general distance measures is the Earth Mover’s Distance (EMD). Although EMD has proven its ability to retrieve similar images in an average precision of around 95%, high execution time is its major drawback. Embedding EMD into L₁ is a solution that solves this problem by sacrificing performance; however, it generates a heavily tailed image feature vector. We aimed to reduce the execution time of embedded EMD and increase its performance using three dimension reduction methods: sampling, sketching, and Dimension Reduction in Embedding by Adjustment in Tail (DREAT). Sampling is a method that randomly picks a small fraction of the image features. On the other hand, sketching is a distance estimation method that is based on specific summary statistics. The last method, DREAT, randomly selects an equally distributed fraction of the image features. We tested the methods on handwritten Persian digit images. Our first proposed method, sampling, reduces execution time by sacrificing the recognition performance. The sketching method outperforms sampling in the recognition, but it records higher execution time. The DREAT outperforms sampling and sketching in both the execution time and performance.