| Popis: |
Image processing techniques in consumer smartphones have brought about a revolution in the quality of photographs taken by absolute novice users. Even so, the photographers are hindered by unsuitable imaging conditions in their attempts to capture desirable pictures. One such condition is the presence of grills, fences or enclosures between the camera and the scene of interest. This results in the presence of occlusions in the clicked image, which, if removed, would make the image more usable. Various image processing techniques have been explored to remove the occlusions present in the form of periodic fence structures, both in spatial and frequency domain. Earlier works in this area depended on machine learning algorithms to train a model using each pixel, and qualifying it as a fence pixel or not, in order to refine the fence mask. The obvious drawback of such an approach is that the model would need to be trained on targeted input images as per the type of image and so the solution is not generic. Using a machine learning algorithm is also quite time-consuming. In this paper, the fence is segmented using frequency domain analysis and we have used the connected component labelling method to refine the fence mask, instead of using the machine learning approach. The efficiency and accuracy of the results obtained through this method have been discussed. |
