Lajos Konyha

Threshold procedures and image segmentation

Several fundamental image segmentation methods exist for the calculation of threshold values. The simple solutions provide typically inaccurate results, while the complex procedures require huge computation resources and/or long processing time (for example, OTSUs method for image thresholding in case of determining several levels). One of the easiest procedures is to divide the colors into two domains by defining a threshold level (between the trivial thresholds) in the intensity signal, or, in case of several thresholds (L count including the trivial ones) distinguishing L1 number of regions. The optimal threshold levels are always dependent on the image content, therefore the adaptive methods always provide much better results. Such a procedure is the adaptive threshold (ATH) algorithm (Andreas E. Savakis, 1998) (J. Sauvola et al., 1997), which is easy to implement for grayscale images, and with slight modifications it is also suitable for color pictures

Character localization in video sequences

Character localization in video sequences is considered a complex and diverse problem. This task can be decomposed into two clearly separate problems: localizing the character regions in the source image, and recognizing characters within the regions found. This paper presents a character localization strategy that significantly reduces the computing requirement of this task. First we present the edge detection algorithms considered. Then the results are used to create a histogram of the source image that helps localize character coordinates within the image. Finally, erroneously found regions are discarded using three rapid and effective solutions, featuring very high reliability of the proposed methods.

Implementation of MPEG4-based transmission in DVB-T technology

Implemented in the 1990s, the MPEG-2 standard as one of todays most widely used image compressing methods provides the basis for baseband data transmission in DVB technology. However, the requirements of the current multimedia applications exceed the capabilities of MPEG-2 systems. On the other hand, these requirements are the driving forces of the MPEG-4 related developments worldwide. Our aim is to contribute to these activities by elaborating procedures that are more effective than the image compressing methods used in MPEG-2 systems, as well as by integrating them into DVB systems

Content-based mesh generation algorithm

The work presents an alternative method for encoding shape-information of video objects, which is a new requirement in the recent evolution of video codec scheme: compressing and transmitting object-based video streams. The proposed method takes the advantage of the latest development in video encoding - motion compensation on the base of mesh presentation of frame -to complete its target. The content-based mesh is deployed to represent video objects, which are detected by a previous video object detector (in a consistent case of scene, detection can simply be a segmenting phase of video frame) or from the whole frame. This special mesh structure constructs the base for successive motion compensation based on mesh-vertices. With content-based mesh, shape information of objects is represented by polygons, therefore it is lossy. Increasing the number of boundary vertices at the cost of bandwidth can alleviate the problem. Integrating the proposed method into a full VOP codec scheme is our perspective.

DEPLOYMENT OF CONSTRAINED DELAUNAY MESH IN VOP SHAPE CODING

The work presents an alternative method for encoding shape-information of video objects, which is a new requirement in the recent revolution of video codec scheme: compressing and transmitting object-based video streams. The proposed method takes the advantage of the latest development in the video encoding – motion compensation on the base of mesh presentation of frame – to complete its target. The so-called constrained Delaunay mesh is deployed to represent video objects, which are detected from a previous video object detector (in a consistent case of scene, detector can simply function like a segmenting phase of video frame). This special mesh structure in one hand constructs the base for successive motion compensation taking root on mesh-vertices. In another hand, the virtue of constrained property opens a possibility to encode the shape of video object implicitly, therefore it reduces the amount of transmitted information. With constrained Delaunay mesh, shape information of objects is presented with polygon, therefore lossy. The problem can be alleviated by increasing the number of boundary vertices at the cost of bandwidth. The work also puts a stress on controlling this trade-off. Results of integrating the proposed method into a full VOP codec scheme are encouraging.