Image Compression versus Matching Accuracy

The improvements of aerial image scanners regarding the scan resolution and in particular the development of new digital cameras (e.g. the DMC of Z/I Imaging) comes along with the production of Gigabytes of image data. To handle these big data sets image compression algorithms are applied. The number of developed image compression procedures is growing and investigations of the consequences for photogrammetric image processing are becoming very important for users of compressed image data. This paper investigates theoretically and experimentally the impact of different compression algorithms on image matching quality. The focus is laid on automated point transfer using the area based least squares matching principle. Other automated processes like aerial triangulation, DTM acquisition and relative orientation often directly rely on the results of the employed point transfer process thus consequences on these processes are inferable. In the experimental investigations simulated stereo imagery and real stereo image pairs are taken into account. Representative algorithms of three different compression methods, the JPEG, Wavelet and Fractal compression, are used to verify and quantify the theoretically found relation between matching accuracy and compression ratio. Regarding the peak-signal-to-noise-ratio (PSNR) of the images induced by compression, Wavelet compression proves to be superior to JPEG and Fractal compression. With respect to the success rate and matching accuracy of point transfer a continuous decrease is to be expected with increasing compression rate. But unlike the PSNR results a changed ranking of the compression methods has to be observed. The highest matching success quotas combined with the most precise point transfer were achieved with the JPEG algorithm. Wavelet compression followed quite closely but Fractal compression dropped down significantly, therefore is not recommended for photogrammetric applications. For high precision point transfer, which is required in most photogrammetric applications, the investigations show that compression ratios not lower than 1:5 can be tolerated. Using images with lower compression ratios will produce point transfer errors above 0.1 to 0.2 pixels.