Catálogo de publicaciones - libros

Single–chip color cameras use a color filter array to sample only one color per pixel. The missing information is interpolated with demosaicking algorithms. Several state–of–the–art and more recent demosaicking methods are compared in this paper. The aim is to find the method best suited for use in computer vision tasks. For this, the mean squared error for various images and for typical color spaces (RGB, HSI and Irb) is measured. The high inter–channel correlation model which is widely used to improve interpolation in textured regions is shown to be inaccurate in colored areas. Consequently, a compromise between good texture estimation and good reconstruction in colored areas must be found, e.g. with the methods by Lu et al. and Freeman.

One of the most important factors to create an atmosphere or a global impression of the image seems to be the feature of the whole image as texture since atmosphere is not so much depended on the details of each part of the image. In order to reproduce atmosphere of the landscape images we made model patterns, reconstructing texture of the images. Full-color landscape images were segmented into patches of same color after converting them into 256-color images and analyzed sizes and numbers of color patches. We made six reconstructed model patterns from the data of color patches. Some of the reconstructed texture patterns are similar in global appearance to the original landscape images although details of the original images are completely destroyed and we can not understand what they are at all. Each reconstructed patterns were evaluated by subjects to find out good methods to reproduce atmosphere of the landscape and two methods got high score. The results suggest that global texture of image create atmosphere of landscape image.

This paper presents a method for wavelet processing of mammogram images in order to highlight pathological changes which are important for diagnosing breast cancer. The method was used at the image pre-processing stage in a CAD (Computer Aided Detection) system which is under development. The paper presents the preliminary detection results for pathological areas with and without the method under discussion.

An algorithm for detection of sperm cells in images of live samples of semen was elaborated. The algorithm processes an original image and finds sperm cells positions. Presented solution is much more robust than those used in current . The algorithm will be used in future in the computer system supporting automated semen analysis.

An approach for segmentation of ultrasound images using features extracted by orthogonal wavelet transforms is proposed. These features are used for learning a backpropagation neural network. The result of classification is improved by using a neighbourhood information.

Detection of microcalcifications (MCs) in mammograms for early breast cancer diagnosing is a widely investigated subject. A number of methods have been tried out so far, but obtained results are still not satisfactory. To avoid difficulties with comparisons of our results with others’, we present results obtained on mammograms from the Digital Database for Screening Mammography (DDSM), provided by the University of South Florida. In this study, a novel approach to MCs detection based on mathematical morphology is presented. A combination of methods is used for the detection of MCs. The evaluation of the proposed technique is done using a free-response operating characteristic (FROC). Our results demonstrate that the MCs can be effectively detected by the proposed approach.

This paper presents a novel method of bias removing from the Rice distributed data. This method is based on developing the inverse formula applied to the function determining the expected value of the Rice distribution. Two algorithms for magnitude image denoising in a wavelet domain are developed. These algorithms use the proposed method of bias removal for the scaling coefficients correction. The denoising performance is better in comparison with the method based on processing the squared magnitude images.

Current technology makes possible to record various events of a human live, such as meetings simultaneously with several video cameras. Large amount of data is obtained from the each recorded event. However, a problem with presentation of such data in a suitable way occurs. This paper describes an algorithm that can be used in a compact videos generation from several source video streams according to different aspects and requirements.

When very large image families are involved in query processes, methods of content-based image retrieval must be optimized with a goal function determining a computing complexity. A clustering method which at the image retrieval stage ensure minimal number of comparisons of a query image and images from image database is proposed. Clustering can be fulfilled in feature or signal space. Pointwise set maps are used as the tools to find required partitions.

Form processing has a lot of practical applications particularly for frequently used forms in a populous country. In this paper we propose a fully automatic method for identification and segmentation of frequently used forms. This helps reduce a lot of manual processing thereby reducing the delay and cost of updating the database.