Oliver Rockinger

Image sequence fusion using a shift-invariant wavelet transform

In this paper, we propose a novel approach to the fusion of spatially registered images and image sequences. The fusion method incorporates a shift invariant extension of the discrete wavelet transform, which yields an overcomplete signal representation. The advantage of the proposed method is the improved temporal stability and consistency of the fused sequence compared to other existing fusion methods. We further introduce an information theoretic quality measure based on mutual information to quantify the stability and consistency of the fused image sequence.

Detection and tracking of small targets in aerial image sequences under unknown sensor motion

Automatic Target Recognition is typically based on single frame image processing. In this paper we report about our work in improving ATR performance by the exploitation of image sequences using a combination of target detection and tracking. The proposed detection/tracking system consists of three subsystems: (1) the target detection module which is based on a combination of multiresolution neural network target filters which are combined by a probabilistic belief network; (2) the sensor motion compensation system which generates a dense velocity field over the actual image frame, thus estimating the effect of the unknown sensor platform motion in image coordinates and (3) a multi-target-tracker which associates existing target tracks with new observations. By the hand of real world examples we show that the combined detection/tracking method overcomes the problem of spurious false alarms generated by the single frame target detector.

Learning target masks in infrared linescan imagery

In this paper we propose a neural network based method for the automatic detection of ground targets in airborne infrared linescan imagery. Instead of using a dedicated feature extraction stage followed by a classification procedure, we propose the following three step scheme: In the first step of the recognition process, the input image is decomposed into its pyramid representation, thus obtaining a multiresolution signal representation. At the lowest three levels of the Laplacian pyramid a neural network filter of moderate size is trained to indicate the target location. The last step consists of a fusion process of the several neural network filters to obtain the final result. To perform this fusion we use a belief network to combine the various filter outputs in a statistical meaningful way. In addition, the belief network allows the integration of further knowledge about the image domain. By applying this multiresolution recognition scheme, we obtain a nearly scale- and rotational invariant target recognition with a significantly decreased false alarm rate compared with a single resolution target recognition scheme.