摘要
本文提出了基于子图像特征组合的商标图像检索算法,首先对商标图像进行子图像抽取,然后根据子图像单特征计算图像与目标图像的单特征距离,最后基于多特征组合得到图像相似性度量.用Hu不变矩对基于子图像多特征组合的商标图像检索算法进行实验,用PVR指数作为图像检索性能评价准则.实验表明,相对基于全局图像单特征的检索算法,基于子图像多特征组合的商标图像检索算法具有更出色的检索性能,其检索结果更符合人眼的视觉感受.
Trademark plays an important role in market economy, reflecting the merchandise s quality and the manufacturer's credit standing. Trademark image is a kind of artificial image without the complicated objects and background that are present in nature image. Each part of a trademark image has obvious division to other parts of it and forms the relatively independent impression in human brain. Several disconnected subimages can be extracted from the trademark image, and the subimage features can reflect the local features of the whole image which are the supplement to the global image feature in expressing the image content. In trademark image retrieval the subimage features can be used to achieve a better performance.
In this paper we put forward a trademark image retrieval algorithm based on the combination of multiple subimage features. First, the subimages of each trademark image are extracted according to the connectedness. Then the single feature distance of each trademark image and the object image is calculated according to the subimages feature fusion. We improve the fusion criterions by weighting them with the subimage reliability, and the experiments show that the weighted minimum criterion is optimal. The image similarity measure is calculated by combining multiple feature distances, which should be normalized bacause of their different variance ranges.
Hu moments, with the invariance of scale, translation and rotation, have been applied in trademark image retrieval systems widely. We also use Hu invariants to perform the trademark image retrieval and use the PVR value to evaluate the image retrieval performance. PVR value is derived from the retrieval precision and recall. Experiments show that the trademark image retrieval algorithm based on the combination of multiple subimage features is superior to that based on single global image feature in terms of the retrieval performance and human vision perception.
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2002年第1期14-20,共7页
Pattern Recognition and Artificial Intelligence
