Polyhierarchical Image Structure and Invariant Object Detection

Abstract

The paper is related to interdisciplinary research in cluster analysis of big data and primary data acquisition from a color image for object detection using the optimal piecewise constant image approximations with a really minimized total squared error or RMSE. Image segmentation is performed via pixel clustering. Ward's clustering is considered as the main method for minimizing RMSE. For Ward's method, the variability property is disclosed, which consists in a pronounced dependence of RMSE obtained for a given color number on the calculation algorithm or slight modification of input data. To overcome excessive computational complexity avoiding timeconsuming programming, parallel execution of pixel clustering algorithms is used with simultaneous selection of approximation hierarchies that reach RMSE minimums for the color numbers in a given range. The problem of invariant hierarchical object segmentation regardless of the image content is studied. The solution is provided by adjusting the tuning parameters from the condition of segmentation invariance in the given image content. The experiments are presented.