A transparent and flexible development environment for rapid design of cognitive systems

Nearest Neighborhood Grayscale Operator for Hardware-Efficient Microscale Texture Extraction

“…A C++ implementation of the operator and its modifications has been carried out based on a software tool for image analysis and classification [8]. The software implementation offers two output formats, either feature numbers as single-frame images (used for higher-level image processing) or feature histograms, which can be directly employed for classification purposes.…”

“…Using the feature histogram output mode, a reduced nearest neighbor (RNN) classifier [5,8] has been trained to recognize eye shapes. Figure 6 shows the trainings and test class spaces, left half, respectively, right half, reduced to two dimensions using Sammon's mapping [8]. Axis captions are omitted because they are a nonlinear, adaptive function of the input-feature vector, and would carry little meaning with respect to the original features.…”

“…This example uses a feature histogram vector composed of the 16 vertical/horizontal features shown in Figure 1. An automated feature selection has been employed to single out the features having the most impact on correct classification [5,8], thus improving classification quality and speed, since classifiers trained on very high-dimensional data tend to "learn" the training sample set, while not being able to generalize well. This is also evinced by the fact that if the features produced by the complete N 8 neighborhood are presented to the classifier, its classification quality decreases to 87.3%, evidently not able to cope with generalization in the context of the resultant increase in search space dimensionality.…”

“…For an N 4 neighborhood, all possible operator outcomes and their respective feature numbers are given in Figure 2. As can be seen, a variety of local grayscale texture information is captured by the operator, ranging from single significant points (feature 15), continuous lines (6,9), terminated lines (7,11,13,14), corners (3,5,10,12), T-sections (1,2,4,8) to complete intersections (0).…”

“…While the software implementation is primarily intended as a stand-alone tool in the framework of the system discussed in [8], it has also been used to verify the LOC pixel cell. The software LOC can be operated in an HDR mode with 16 Bit resolution, which has been used to show validity of the modified LOC concept, especially the new significance threshold introduced in (5), in an HDR context.…”