Integrated Object Detection and Deconvolution

A vision model is a framework for carrying out object detection. (In a context other than astronomy, feature detection may be a more appropriate term.) A vision model may be simple e.g. background + object(s) superimposed, or it may be more sophisticated e.g. involving a multiscale transform. The benefits of considering carefully the vision model used, for very fine-tuned work, have been discussed in Bijaoui and Rué (1995), Rué and Bijaoui (1997), and Starck et al. (1998).

If we know that deconvolution will improve the result, then we have to integrate deconvolution with our vision model. An integrated algorithm for doing this is described in Starck (1999).

Fig. 1 (compressed Postscript, 180kB) shows isophotes of objects detected using a vision model - multiscale transform with noise modeling - without deconvolution. ISOCAM (a detector on board the ISO satellite) was used, and the data collected using the 6 arcsec lens at 6.75 microns. This was a raster observation with 10s integration time, 16 raster positions, and 25 frames per raster position. The non-stationary noise was modeled using a root-mean-square error map (see Starck et al., 1999). In the Fig., the isophotes are overplotted on an optical image, NTT, band V (from ESO's Paranal observatory), in order to identify the infrared sources. Fig. 2 (compressed Postscript, 173kB) shows the same result, but this time based on vision modeling with deconvolution. The objects are the same, but the photometry is improved. It is clearly easier to identify the optical counterparts of the infrared sources.