Seite - 124 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Fig. 4: Gabor filter composition: (a) 2D sinusoid oriented at 30◦ with the x-axis, (b) a Gaussian kernel, (c) the corresponding Gabor filter [6]. were used to generate covariance descriptors of Gabor filter outputs with one frequency and six orientations. Nearest neighbor classification was used for evaluation. Figure 5 shows a whole steel block and the segmentation results. From Figure 5b it is obvious that the classification output delivers plausible results for the trained type of steel, although the trans-crystalline areas are not perfectly classified if the orientation of the solidification structure does not perfectly match the trained ground truth data. B. Spatial Filter Bank The paper presented by Ahmadvand and Daliri [1] intro- duces a way to perform invariant texture classification by using a spatial filter bank in multi-resolution analysis. The generated features comprise l1-norm, standard deviation and entropy calculated from the spatial filter bank results of the original patch and the discrete wavelet transformed patch. Proposed filters are Gaussian, Laplacian of Gaussian and local standard deviation. Same as for Gabor filters, two different patch classes are used to set up two feature matrices. For classification simple Mahalanobis distances between the feature vector and the matrices are calculated to determine class affiliation. Although certain regions (middle and bottom in Figure 5c) are extracted more homogeneously than in the Gabor filter approach, the classification output does not yield a satisfactory result as it is too dependent on selection of training patches. The filter bank matches good within direct surroundings of training patch areas, whereas other areas cannot clearly be separated. (a) Original image. (b) Gabor filter output with nearest neighbor classification. (c) Spatial filter bank output with Mahalanobis distance classification. Fig. 5: Original image and segmentation output. C. Local Binary Patterns (LBP) LBP [10] are used to describe the surrounding of a pixel. This is done by comparing a pixel to each of its neighbors (which [10] defines by radius and number of points on the consequential circle). Given eight neighbors LBP result in an eight digit binary number where each digit gives information about whether the center point value is greater/equal or smaller than its neighbor. To retrieve information about a larger area LBP for each pixel in that area are summed up in a histogram illustrated in Figure 6. Fig. 6: LBP histogram generation. To be able to determine certain edge and line infor- mation of an area’s histogram, we decided to summarize inverted patterns, same orientation patterns or patterns that just describe noise. Overall dominating bins like noise and white/black dots are deleted from the histogram. Following those steps, it is possible to determine features (histogram bins) that correlate with the desired regions. 124