Chinese calligraphy is a very special style of handwriting and direct character recognition is very difficult. Content-based keyword spotting is more feasible than recognition-based retrieval for calligraphy document. In this paper,we propose a novel Elastic Histogram of Oriented Gradient( EHOG) descriptor for calligraphy word spotting. The presented feature is a modification of Histogram of Oriented Gradient( HOG), widely used in human detection. In our approach,the input word image is partitioned into non-uniform rectangular cells according to the calligraphy character pixel intensity,and then in each cell a histogram of orientation is accumulated dynamically. Moreover,we adopt Derivative Dynamic Time Warping( DDTW) for image feature matching,which achieves good performance in gesture recognition. Experiments demonstrate a very significant improvement when comparing our proposed feature with previously developed ones,and also show DDTW produces superior alignments between two calligraphy character feature series than DTW.
