Detecting Pills in Fabric Images Based on Multi-scale Matched Filtering
Detecting Pills in Fabric Images Based on Multi-scale Matched Filtering
复制标题
DOI：
10.1177/0040517508099913
复制
发表时间：
2009-10
影响因子：
2.3
通讯作者：
Xia Chen;Zengbo Xu;Ting Chen;Jun Wang;Liqing Li
中科院分区：
文献类型：
--
作者：
Xia Chen;Zengbo Xu;Ting Chen;Jun Wang;Liqing Li
介绍了基于多尺度匹配滤波的织物图像中毛球的识别与分割方法。根据药丸的渐近强度变化，采用高斯函数对药丸进行建模。设计了一种匹配过滤器，用于检测织物颗粒。然而，大药丸和小药丸可能在图像中随机分布，很难与滤波器同时检测到。为了解决这一问题，利用小波函数的多尺度展开和收缩特性，生成一组具有可变大小卷积核的匹配滤波器，并与织物图像进行卷积。然后在匹配滤波的每个输出处识别小球，通过合并多尺度匹配滤波的所有输出来确定小球是否存在于局部区域。最后，利用基于均值和加权方差的自适应阈值方法，对已经确认存在的局部区域的药丸进行进一步分割。实际图像上的丸剂检测结果表明，该方法可以满足不同大小丸剂的识别和分割，有助于定量描述丸剂信息。
The recognition and segmentation of pills in fabric images based on multi-scale matched filtering are introduced in this paper. According to the asymptotic intensity change of a pill, the pill was modeled using a Gaussian function. A matched filter is designed for detecting fabric pills. However, big pills and small pills may distribute randomly in an image and are hard to be detected simultaneously with the filter. To solve this problem, a group of matched filters, with variable-size convolution kernels using the multi-scale expansion and shrinkage properties of the wavelet function, are generated and convolved with the fabric image. Then pills, at each output of matched filtering, are recognized and we determined whether pills exist in the local area by merging all of the outputs of the multi-scale matched filtering. Finally, pills in the local area, where the existence of pills has been confirmed, are further segmented using an adaptive threshold method based on mean and weighted variance. The detection of pills on a real image shows that this method can satisfy the recognition and segmentation of different size pills and is helpful for quantitatively describing the pilling information.