A proposed multi-scale approach with automatic scale selection for image change detection

In this paper a system for multi-scale change detection with automatic scale selection is proposed. The generation of the multi-scale data set is performed based on the fractal net evolution approach (FNEA). The set of scales used by FNEA are optimally selected from the scale domain ensuring that the selected levels present a good enough representation of the scale domain. A pattern search module is used to select good enough set of scales with the least redundancy. The change detection is performed on each scale individually. For each individual object in a specific scale change indicators are extracted for the pixels corresponding to this object in the base-scale images. After extracting the change indicators for each scale, the extracted indicators are thresholded to obtain a per-scale binary change map. To obtain the final change map, a scale-driven fusion of all the extracted change maps is performed. The fusion is based on detecting for each pixel the preferred scale to obtain its change information. The best scale for an object is the scale where the object area keeps static/almost static while moving from one scale to the next scale(s). The proposed system proves advantageous over other change detection systems.