Shift-Invariant Motif Discovery in Image Processing 'Best Paper Award'

Nowadays, the boost of optical imaging technologies results in more data with a faster rate are being collected. Consequently, data and knowledge discovery science has become an attractive and a fast growing topic in several industry and research area. Motif discovery in image processing aims to tackle the problem of deriving structures or detecting regularities in image databases. Most of the motif discovery methods first convert images into time series and then attempt to find motifs in such data. This might lead to information loss and also the problem of inability to detect shifted and multi-scale image motifs of different size. Here, a method is proposed to find image motifs of different size in image datasets by applying images in original dimension without converting them to time series. Images are inspected by the Complex Quad Tree Wavelet Packet transform which provides broad frequency analysis of an image in various scales. Next, features are extracted from the wavelet coefficients. Finally, image motifs are detected by measuring the similarity of the features. The performance of the proposed method is demonstrated on a dataset with images from diverse applications, such as hand gesture, text recognition, leaf and plant identification, etc.