Author: Božidar Potočnik
Mentor: Prof. Dr. Damjan Zazula
Co mentor: Prof. Dr. Franc Solina
Date: Sept. 14, 2000
Prediction-Based Object Recognition from a
Sequence of Images
Keywords: digital image processing, object recognition, image sequence, moving object, object tracking, segmentation, region growing, predictor-corrector methods, Kalman filter, prediction, modelling, ultrasound images, speckle noise, ovarian follicles
Abstract: In doctoral thesis, automated object recognition from a sequence of 2D images is dealt with. We suppose that no prior knowledge about the sequence is disposable, with the exception of type of images and objects that are searched for, neither is possible to assure that variations from image to image are small. All objects in the image are tracked, each object, however, could move with its own motion. The thesis begins with an overview of existing methods and techniques for processing the sequence of digital images, emphasising the recognition methods. The methods are critically assessed and are classified into five areas: a) methods for motion analysis in images, b) segmentation methods based on motion, c) active contours, d) techniques for 3D images, and e) predictor-corrector methods.
The research is continued with construction of two new recognition systems for automated object recognition from a sequence of 2D images. To be able to
verify the recognition efficiency easier, real ovarian ultrasound images with follicles and also artificial images are tested. A new procedure for creating artificial images is developed. A 3D simulation ovarian model with follicles is formed based on oval geometric constructs. The model is calibrated; afterwards, a computer simulator for projecting the cross-sections through ovarian model at an arbitrary angle and shift is set.
Basic procedure for recognition of dark oval objects in a static 2D image, i.e. an algorithm for static 2D images, follows. Beside the region growing used as a segmentation method, a weighted gradient and a set of statistical tests are introduced for recognition. The developed algorithm is verified using artificial and real ovarian ultrasound images. In the artificial images, all conditions are precisely known, while for real images the efficiency of method is measured according to the readings given by a doctor specialists. Appropriate quantitative measures for assessing the efficiency of algorithm are described. Recognition rate for the artificial images is around 90 % and for real images around 78 %. An average absolute distance between the boundary of recognized and correct follicles is in both cases about 1 mm.
Then we pass from object recognition in a single static 2D image to the recognition where entire information from an image sequence is considered. Through accurate mathematical formulation of a problem we prove that it is the most optimal if recognition history from previous images is built in a Kalman filter and is used for recognition on current image. The new algorithm for object recognition from an image sequence with prediction procedures, i.e. prediction algorithm, is described in detail. It is based on the Kalman filter. The measurement system is realized with the algorithm for static 2D images. Based on measurements in the first image, an object model is set. Afterwards, this model is modified from image to image with Kalman filter regarding new measurements. The model calculated for particular image defines the new best estimate of the object searched for. Possible problems appearing at such a recognition are identified, and beside the basic prediction algorithm three improvements are given: for the case of big errors in measurements, for considering the locality of object contours, and for correcting the unsuitable measurements. The final prediction algorithm is given in a pseudocode form.
The prediction algorithm is tested on sequences of artificial and real ovarian ultrasound images with follicles. For the assessment, already introduced quantitative measures are used. The obtained results are much more compact and accurate with the prediction algorithm than with the algorithm for static 2D images. The number of misidentified follicles is considerable lowered (also to 80 % according to starting values). It points out that with such processing of image sequences reliability of the obtained results is greatly increased. The prediction algorithm in some difficult situations detects follicles even more correctly than a doctor expert does.