变形网格及其在图像识别中的应用

Application of Deformable Grid in Image Recognition

网格特征是图像识别中一类重要特征,而变形模板对于复杂的图像识别问题如字符识别、数字识别、图标识别等表现出很好的性能,但变形模板很费时.该文针对网格特征提出了变形网格,并分析了变形模板与变形网格之间的近似等价性.所提方法对网格进行变形而不是对图像变形,因而速度比变形模板快得多,而且性能相差不大.把这种方法分别应用于图标识别和脱机手写汉字识别.图标识别实验中变形模板使识别率提高了7.5％,而变形网格使识别率提高了7.3％.手写汉字识别实验中变形模板使识别率提高了6.1％,而变形网格使识别率提高了5.8％.考虑到变形网格比变形模板快得多,所以这种方法是有优势的.

Grid feature is a popular feature extraction scheme in image recognition, and usually higher performance could be obtained with the help of deformable template, especially in such difficult image recognition tasks as character recognition, digit recognition, logo recognition etc. The critical shortcoming of deformable template is time-consuming. In this paper the method of deformable grid is proposed to compensate for the shortcoming of deformable template. Firstly certain grid must be superimposed on image, then unlike deformable template where deformation is applied to image, various deformations are applied to grid. Because the number of grid is much less than that of pixels in image, the method is much more timesaving comparing to deformable template. The approximate equality of deformable template and deformable gird is also analyzed. The method is evaluated by two image recognition experiments, namely, logo recognition and off-line Chinese character recognition. The improvements in recognition rate by 7.3% in first experiment, and 5.8% in second one are obtained by the use of deformable grid.