Abstract: A novel image texture feature descriptor is proposed which is insensitive to Gaussian blur and Gaussian noise.The four scales horizontal high-pass components and vertical high-pass components of image are extracted by dyadic wavelet transform.These components are arranged at each pixel in image，and then uniformly quantized in an eight-dimensional space，represented as an integer using binary coding.At last，a histogram sequence is computed by concatenated histogram from each region which is the descriptor of the image.Due only using the high-frequency coefficients，the descriptor is more robust to Gaussian noise.Experiment on face image recognition shows that the method has attractive features than LBP and LPQ which is the state-of-art local feature descriptors.

Key words: dyadic wavelet, quantization feature, histogram feature, face recognition

摘要： 提出了一种对高斯模糊和高斯噪声鲁棒的图像纹理特征描述算子。通过对图像进行二进小波变换，得到图像4个尺度的横向和纵向滤波系数，在每个像素点排列这些系数，然后量化为一个8维的空间，并通过二进制编码变成一个整数。然后对图像进行分区域统计得到一个直方图序列作为图像的特征描述算子。由于仅利用了图像的边缘等高频信息，所以对高斯噪声更具有鲁棒性。通过人脸图片识别实验表明特征描述算子具有比LBP、LPQ等著名的特征描述算子具有更好的性能。

关键词: 二进小波, 量化特征, 直方图特征, 人脸识别