Image processing

Panoramic image from video

The task of constructing panoramic image from video address the following sub-tasks: control points selection; model linking algorithm for image; building panoramas, based on the linked images. To select feature points on the image modified Harris detector was used. To determine the degree of proximity of feature points Euclidean metric was used. For the dropout rate of false matches between the feature points of the approach used by RANSAC. Weighted sum of overlapped images was used to avoid boundary effects.

Source video:

Result

Source video:

Result

Detailed examples

Feature detection

Image binding

Image binding with RANSAC

Image quality improvement

Retinex approach is based on the removal of the local inhomogeneity of illumination of objects. Poorly illuminated objects become more apparent, their details are restored. As part of our research the following tasks were solved:

• Comparison of different versions of RETINEX algorithm;
• Comparison of results of work in different color spaces;
• Possibilities of optimization algorithm have been investigated.

More examples:

Deblurring

The problem of restoring images distorted by some arbitrary function. Special cases of this effect — blurred and out of focus images. That effect can be obtained with the shooting of fast moving vehicle or by rotating the camera around its axis. The following approaches was for reconstruction: one-pass Wiener filtering, iterative algorithms (Van Cittert, Landweber, Richardson-Lucy), the method of "blind" reconstruction (image blind deconvolution). The best results were obtained for cases where there is a priori information about the nature of the pixel spear function. Experiments have been conducted with different lengths smear (1 ~ 200 pixels). Modified Wiener filtering with noise suppression used to restore. In addition, original approach was disigned to supress false contours in the reconstructed image.