Classification of Rotated and Scaled Textured Images

大多数纹理分类方法都是在假设测试图像与训练图像具有相同的方向和尺度的前提下提出分类问题的。因此，即使测试图像的尺度方向与训练图像略有不同，它们的表现也很差。本课题研究了一种基于光谱矩提取的旋转尺度不变量的纹理分类新方法。许多研究者研究了由矩导出的旋转不变量。然而，使用它们进行模式识别的成功结果仅限于具有有限大小对象的二值图像。由于灰度图像的边界上存在非零值，因此在灰度图像中应用不变矩是很困难的。利用谱密度函数（SDF）的矩来代替图像的灰度矩来获得旋转不变量，因为SDF具有很好的特性。当图像被旋转和缩放时，它的SDF旋转相同的角度并被反向缩放；此外，大多数纹理图像的光谱功率集中在原点附近，并向高频方向迅速衰减。此外，我们还设计了稳定的旋转矩；我们将计算力矩的积分区域定义为圆，并通过在圆中设置图像总功率的恒定速率来调整其半径以适应变化。由于谱矩对尺度有一个简单的性质，我们可以用旋转不变量的比值来定义RS不变量。我们用来自Brodatz专辑的16种纹理进行了实验，验证了该方法的有效性，并取得了很高的分类率。我们将该技术应用于检测(1)纹理图像的旋转角度和比例因子，(2)衣服缺陷，(3)乳房x线照片中的微钙化，并取得了良好的效果。

Most of texture classification approaches pose classification problems under the assumption that a test image possesses the same orientation and scale as training images. According1y, they perform poorly even though the orientation of the scale of the test image differs a little from those of the training ones.This project studies a new technique for texture classification based on rotation and scale (RS) invariants obtained from spectral moments. Rotation invariants derived from moments have been studied by many researchers. However, successful results for pattern recognition using them are limited only to binary images with a finite-sized object. Application of the invariant moments to gray-scale images is difficult because the images have non-zero values on the image boundaries.Instead of the gray-scale moments of the image, we use the moments of its spectral density function (SDF) to obtain rotation invariants because the SDF has advantageous properties. When an image is rotated and scaled, its SDF rotates by the same ang1e and is inversely scaled ; moreover, the spectral powers of most textured images are concentrated around the origin and decay rapidly toward high frequencies. In addition, we scheme to obtain stable rotation moments ; we define as a circle the integral region for moment calculation and adjust its radius to sca1e change by setting a constant rate of the image total power in the circle. Since spectral moments have a simple property with respect to scale, we can define RS invariants by a ratio of rotation invariants.We confirmed the validity of the proposed method in experiments using 16 textures from Brodatz album with very high classification rates. We applied this technique to detection of (1) the rotation ang1e and the scaling factor of textured images, (2) the defect of c1othes and (3) the micro calcifications in mammograms and obtained prosperous results.