a Novel Similarity Assessment for Remote Sensing Images via Fast Association Rule Mining

Similarity assessment is the fundamentally important to various remote sensing applications such as image classification, image retrieval and so on. The objective of similarity assessment is to automatically distinguish differences between images and identify the contents of an image. Unlike the existing feature-based or object-based methods, we concern more about the deep level pattern of image content. The association rule mining is capable to find out the potential patterns of image, hence in this paper, a fast association rule mining algorithm is proposed and the similarity is represented by rules. More specifically, the proposed approach consist of the following steps: firstly, the gray level of image is compressed using linear segmentation to avoid interference of details and reduce the computation amount; then the compressed gray values between pixels are collected to generate the transaction sets which are transformed into the proposed multi-dimension data cube structure; the association rules are then fast mined based on multi-dimension data cube; finally the mined rules are represented as a vector and similarity assessment is achieved by vector comparison using first order approximation of Kullback-Leibler divergence. Experimental results indicate that the proposed fast association rule mining algorithm is more effective than the widely used Apriori method. The remote sensing image retrieval experiments using various images for example, QuickBird, WorldView-2, based on the existing and proposed similarity assessment show that the proposed method can provide higher retrieval precision.