Abstract :
A new unified approach to object and change detection is presented that involves clustering and analyzing the distribution of pixel values within clusters over one or more images. Cluster-based anomaly detection (CBAD) can detect man-made objects that are: (1) present in a single multiband image; (2) appear or disappear between two images acquired at different times; or (3) manifest themselves as spectral differences between two sets of bands acquired at the same time. Based on a Gaussian mixture model, CBAD offers an alternative to compute-intensive, sliding-window algorithms like Reed and Yu´s RX-algorithm for single-image object detection. It assumes that background pixel values within clusters can be modeled as Gaussian distributions about mean values that vary cluster-to-cluster and that anomalies (man-made objects) have values that deviate significantly from the distribution of the cluster. This model is valid in situations where the frequency of occurrence of man-made objects is low compared to the background so that they do not form distinct clusters, but are instead split up among multiple background clusters. CBAD estimates background statistics over clusters, not sliding windows, and so can detect objects of any size or shape. This provides the flexibility of filtering detections at the object level. Examples show the ability to detect small compact objects such as vehicles as well as large, spatially extended features (e.g., built-up and bomb-damaged areas). Unlike previous approaches to change detection, which compare pixels, vectors, features, or objects, cluster-based change detection involves no direct comparison of images. In fact, it is identical to the object detection algorithm, different only in the way it is applied. Preliminary results show cluster-based change detection is less sensitive to image misregistration errors than global change detection. The same cluster-based algorithm can also be used for cross-spectral anomaly detection. An example showing the detection of thermal anomalies in Landsat Thematic Mapper imagery is provided.
Keywords :
Gaussian processes; data acquisition; geophysical signal processing; image registration; image representation; image sequences; object detection; pattern clustering; terrain mapping; vector quantisation; Gaussian distributions; Gaussian mixture model; Gaussian processes; Landsat Thematic Mapper imagery; background pixel values; background statistics; cluster-based anomaly detection; compute-intensive sliding-window algorithms; cross-spectral anomaly detection; filtering detections; global change detection; image acquisition; image misregistration errors; image representations; image sequence analysis; imagery change detection; man-made object detection; object occurrence frequency; pattern clustering; pixel value distribution; single multiband image; single-image object detection; spectral differences; thermal anomaly detection; vector quantization; Change detection algorithms; Clustering algorithms; Filtering; Frequency; Gaussian distribution; Image analysis; Object detection; Pixel; Shape; Statistical distributions;
