基于自动定位分割的图书识别框架

A Framework for Book Cover Recognition Based on Automatic Location and Segmentation

提出一种基于自动定位分割的图书识别算法,主要包括对拍摄图像进行图书封面区域的自动定位、感兴趣区域(ROI)分割、标准形状矫正以及有效区域的特征提取与相似性匹配部分.自动定位分割部分根据图书封面的几何形状特点,通过基于霍夫变换的形状检测算法对自然场景下拍摄的图书封面图像进行有效区域定位,对ROI进行分割并根据逆仿射变换将其矫正到标准形状;然后对获取的有效图书封面区域进行基于改进的尺度不变特征变换(SIFT)的特征点检测和特征描述,并采用词包(BOW)方法对其进行特征量化和码本学习,从而将定位分割出的图书图像与数据库中图书源图像进行相似性匹配.实验结果表明,对包含一定复杂程度背景的图书图像进行准确的定位分割和矫正,在很大程度上影响着基于特征匹配的图书识别技术的精确度.

This paper put forward a framework for book recognition based on automatic location and segmentation, which includes the stage of book automatic location, ROI segmentation, standard shape correction, effective feature extraction and similarity matching. Automatic location and segmentation stage based on the geometric shape characteristic of book, takes effective regional orientation and ROI segmentation through the hough transform algorithm , and then correct the real book cover areas to a standard shape according to the inverse affine transformation. The stage of feature extraction and similarity matching mainly used the improved SIFT algorithm to conduct feature points detection and generate feature description, and then conduct the similarity matching by the popular BOW method which generates codebook and quantizes the features. The results of our experiments show that the accurate location and segmentation of the book cover from a image which contain some appropriate background and taking standard shape correction have a significant impact on the precision of recognition.