|
Random and systematic radiometric distortion in the images sequence is a common issue in digital photogrammetric close range applications. A digital image filtering enables correction of blur and elimination of noise on digital images. The paper analyses the influence of digital filtering of blurred and noisy close-range images on the subpixel accuracy of signalized points measurement and accuracy block bundle adjustment. Blurred (out of focus) images were processed with 6 linear high-pass filters, whereas 5 linear low-pass filters and 1 nonlinear median filter were applied for noisy images. The photogrammetric measurement was conducted on test field in Institute of Photogrammetry and Remote Sensing, Dresden University of Technology. The test field consisted of the total of 220 signalized, retro-reflective points (65 coded points, 14 bit code). The test field was registered on eleven convergent and normal color photos from the distance of ca YF = 5 m using digital SLR Kodak DCS Pro 14n (resolution 4500.3000) camera. Adobe Photoshop CS6 and Corel PHOTO-PAINT X6 software were used for generation of blurred images (Gaussian blur, radius r = 2 and radius r = 3) and noisy images (random Gaussian noise, level = 25, density = 50). The fully automatic points measurement on the digital images using center weighted method, the bundle adjustment including self-calibration with additional parameters for modeling systematic imaging errors was determined in the AICON 3D Studio software package (AICON 3D Systems GmbH, Germany). The optimal correction terms contain additional parameters for the compensation of radialsymmetric A1, A2 lens distortion and radial-asymmetric tangential B1, B2 lens distortion as well as affinity C1and shear of the digital sensor coordinate system C2. High-pass filtering of blurred digital images (blur r = 2) is essential for measurement of structural signalized points using center weighted method. In the case of significantly blurred images (blur r = 3) the typical Laplace filters do not sharpen the images to the extend enabling measurement. Laplacian of Gauss (mask 5×5 pixel) turned out to be the only efficient high-pass filter. The low-pass filtering and nonlinear median filter of noisy digital images does not influence the measurement accuracy and values of adjusted parameters. The center weighted operator is robust in the structural signalized points measurement on the insignificantly blurred images and is resistant to occurrences of random noises. In all of the analyzed variants Sigma 0 mean value after the combined bundle adjustment with the self-calibration amounted to .0 = 0.07 pixel (sensor pixel size p.HV = 7.9 µm). The RMS image coordinates residuals after adjustment amounted to RMS Vx' = 0.05 pixel and RMS Vy' = 0.06 pixel and the RMS standard deviation of calculated object coordinates respectively amounted to: RMS SX = 0.05 mm, RMS SY = 0.1 mm, RMS SZ = 0.04 mm.
|