Picard R W, Gorkani M
Perceptual Computing Section, MIT Media Laboratory, Cambridge, MA 02139.
Spat Vis. 1994;8(2):221-53. doi: 10.1163/156856894x00341.
An algorithm for detecting orientation in texture is developed and compared with results of humans detecting orientation in the same textures. The algorithm is based on the steerable filters of Freeman and Adelson (IEEE Trans. PAMI 13, 891-906, 1991), orientation-selective filters derived from derivatives of Gaussians. The filters are applied over multiple scales and their outputs non-linearly contrast-normalized. The data for humans were collected from forty subjects who were asked to identify 'the minimum number of dominant orientations' they perceived, and the 'strength' with which they perceived each orientation. Test data consisted of 111 grey-level images of natural textures taken from the Brodatz album, a standard collection used in computer vision and image processing. Results show that the computer and humans chose at least one of the same dominant orientations on 95 of the natural textures. Of these textures, 74 were also in 100% agreement on the location of all the dominant orientations chosen by both humans and computer. Disagreements are analyzed and possible causes are discussed. Some apparent limitations in the current filter shapes and sizes are illustrated, as well as some (surprisingly small) effects believed to be caused by semantic recognition and gestalt grouping.
开发了一种用于检测纹理方向的算法，并将其与人类在相同纹理中检测方向的结果进行比较。该算法基于弗里曼和阿德尔森的可控滤波器（《IEEE模式分析与机器智能汇刊》13卷，第891 - 906页，1991年），即从高斯导数导出的方向选择性滤波器。这些滤波器在多个尺度上应用，其输出进行非线性对比度归一化处理。人类的数据是从40名受试者那里收集的，他们被要求识别自己感知到的“主要方向的最小数量”以及感知每个方向的“强度”。测试数据由111张取自布罗达茨相册的自然纹理灰度图像组成，该相册是计算机视觉和图像处理中常用的标准图集。结果表明，在111张自然纹理中，计算机和人类在95张纹理上至少选择了一个相同的主要方向。在这些纹理中，74张在人类和计算机所选的所有主要方向的位置上也完全一致。对分歧进行了分析并讨论了可能的原因。文中说明了当前滤波器形状和尺寸存在的一些明显局限性，以及一些据信由语义识别和格式塔分组导致的（出人意料地小的）影响。
