目的 为了提取零件表面图像的纹理特征并对其表面粗糙度分类识别,有效提高识别的正确率,提出了联合Gabor小波和改进局部二值模式(LBP)的纹理特征提取方法。方法 针对传统LBP算子忽略了邻域内灰度差幅值特征的问题,提出了M_LBP(magnitude considered LBP)算子。采用Gabor小波对零件表面图像滤波,并计算各子图像 Gabor幅值特征GMM(Gabor magnitude maps)。应用M_LBP算子计算各GMM的M_LBP特征谱,进而构造得到零件表面图像的纹理特征向量,最后通过KNN(K-nearest neighbor)算法对零件粗糙度分类识别。结果 本文提出的算法有效细化了表面图像纹理特征,对粗糙度差别为0.2 μm的零件识别准确率达到98%,远高于利用传统LBP算子提取的纹理信息的识别准确率。结论 本文提出了一种有效细化LBP纹理特征的M_LBP算子,并通过与Gabor小波的结合,突破了传统LBP算子尺度、方向单一,幅值信息被忽略的局限性,能实现较高精度的粗糙度识别。

Objective A method for textural feature extraction based on Gabor wavelet and improved local binary pattern (LBP) is proposed to classify and recognize the surface roughness of a work piece through its image and improve accuracy. Method Given that the LBP operator ignores the magnitude differences between neighbors, the magnitude-considered LBP (M_LBP) operator is proposed. The magnitude of the gray level differences between neighbors is defined as D_c. The gray mean of the image is used as threshold for the binarization of D_c. The binarization result is appended to the top digit of the LBP in this neighbor, which is obtained by dividing the LBP according to the D_c value. Before the recognition of the surface roughness of the work piece, the surface image is obtained with a stereomicroscope and then preprocessed. A self-similar Gabor wavelet filter bank is acquired by changing the scale and orientation parameters. The filter bank is used for surface image filtering. The multi-scale and multi-resolution Gabor texture features of the image are acquired. Afterward, the magnitude of the Gabor texture features is calculated, and the Gabor magnitude maps (GMMs) are extracted. The proposed M_LBP operator is then applied in the GMMs to extract M_LBP feature maps. Based on these M_LBP feature maps, the texture feature vector for each surface image can be constructed. After extracting these texture feature vectors, the k-nearest neighbor (KNN) algorithm is used in roughness recognition. The feature vectors of both the training samples and the samples to be recognized are extracted. Then, the, K nearest neighbors of the samples to be recognized are selected from the training samples. The roughness class of the sample to be recognized can be acquired according to the roughness classes of these K nearest neighbors. Result Different values are selected for (P, R) in the M_LBP operator and for K in the KNN algorithm. The comparative experiment shows that the recognition accuracy is highest when (P, R) is set to (8, 8) and K is set to 4. We use the LBP, Gabor combined with LBP, and Gabor combined with M_LBP to extract texture feature vectors. Through these vectors, we compare the time consumption (0.2886, 0.9546, and 1.1562 s) and the recognition accuracy (74%, 82%, and 98%). The experimental results demonstrate that the proposed method can recognize the surface roughness of the work piece with 98% accuracy and a difference of 0.2 μm, which is better than that of the other two algorithms. Conclusion The proposed M_LBP operator can refine LBP information, and the Gabor wavelet combined with M_LBP overcomes the limitations of LBP, which includes single scale, single orientation, and disregard for magnitude. Hence, the method can be applied in roughness recognition with high precision.