Božidar Potočnik

Automated analysis of a sequence of ovarian ultrasound images. Part I: segmentation of single 2D images

Abstract An improved algorithm is presented for ovarian follicle detection in ultrasound images. This fully automated recognition algorithm is composed of three successive steps. First, initial homogeneous regions are determined. Then, these initial regions are grown. The growing is controlled by average grey-level and by a newly introduced weighted image gradient. In the last stage, those regions are extracted that probably correspond to the follicles. The algorithm has been tested on 50 ovarian ultrasound images. The recognition rate of follicles using this procedure was around 78%. A possible extension of the algorithm deals with the entire information in the ultrasound image sequence, which is covered in Part II of this paper.

Automated analysis of a sequence of ovarian ultrasound images. Part II: prediction-based object recognition from a sequence of images

Abstract Part I of this paper introduced a new algorithm for object detection on a single static image from an image sequence. Part II extends this basic 2D recognition scheme by incorporating knowledge about previous image recognition. A new algorithm is presented for object recognition from an image sequence using prediction procedures. It is based on the Kalman filter (KF). The measurement system is realised with an algorithm for static 2D images. An object model is set based on measurements in the first image of sequence. This model is modified from image to image using the KF in regard to new measurements. The calculation for a particular image defines a new best estimate of the object searched for. This prediction algorithm (PA) was tested on sequences of ovarian ultrasound images with follicles. The obtained results are much more compact and accurate using the PA than with the 2D algorithm only (up to 30% according to the initial values). The number of misidentified follicles is considerably lower (up to 75%).

A hybrid differential evolution for optimal multilevel image thresholding

Real test images were used to perform thresholding using Otsus method with a high level of thresholds.The proposed hybrid is compared with DE, jDE, PSO, ABC, and CS.Algorithms are compared based on PSNR and SSIM metrics.Friedman and Wilcoxon statistical tests are used to show the performances.Proposed hjDE shows superior performance in the quality of the results. Image thresholding is a process for separating interesting objects within an image from their background. An optimal thresholds selection can be regarded as a single objective optimization problem, where obtaining a solution can be computationally expensive and time-consuming, especially when the number of thresholds increases greatly. This paper proposes a novel hybrid differential evolution algorithm for selecting the optimal threshold values for a given gray-level input image, using the criterion defined by Otsu. The hybridization is done by adding a reset strategy, adopted from the Cuckoo Search, within the evolutionary loop of differential evolution. Additionally a study of different evolutionary or swarm-based intelligence algorithms for the purpose of thresholding, with a higher number of thresholds was performed, since many real-world applications require more than just a few thresholds for further processing. Experiments were performed on eleven real world images. The efficiency of the hybrid was compared to the cuckoo search and self-adaptive differential evolution, the original differential evolution, particle swarm optimization, and artificial bee colony where the results showed the superiority of the hybrid in terms of better segmentation results with the increased number of thresholds. Since the proposed method needs only two parameters adjusted, it is by far a better choice for real-life applications.

Multi-Objective Differential Evolution for feature selection in Facial Expression Recognition systems

Feature selection using Multi-Objective Differential Evolution (DEMO) is proposed.This efficient selection is integrated in advanced Facial Emotion Recognition system.Emotion-specific and more discriminative features over all emotions selection strategies.Emotion recognition accuracy of proposed algorithm comparable to state-of-the-art.Feature selection by using DEMO enormously reduces a feature vector length. This paper proposes an efficient feature selection system applied to a Facial Expression Recognition (FER) system. This system, capable of recognizing seven prototypical emotions including neutral expression, is based on a histogram of oriented gradient descriptor (HOG) and difference feature vectors. The emotion feature selection was carried out by using an appropriately modified multi-objective differential evolution algorithm. The number of used features was minimized, while the emotion recognition accuracy of the support vector machine classifiers was maximized simultaneously. The emotion-specific features and the more discriminative features over all emotions selection strategies were developed, whereby the latter strategy proved to be more efficient using the Friedman statistical test. This person-independent FER system with proposed feature selection was validated on three commonly used evaluation databases, where the mean emotion recognition rate was 98.37% on the Cohn Kanade database, 92.75% on the JAFFE database, and 84.07% on the MMI database, while the number of used features lowered up to 89% with respect to the original difference feature vector length. Compared to the state-of-the-art, the proposed FER method offers good results, while also greatly lowering the number of used features, which, in return, minimizes the computational cost of training the classifiers. The optimization proposed in this paper can be generalized easily to a feature selection for an arbitrary multi-objective, as well as many-objective, problem.

Estimating correspondence between arbitrarily selected points in two widely-separated views

When monitoring events on a building site using a system of multiple cameras, it is necessary to establish correspondences between the acquired imaging material. The basic problem when attempting this task is the establishment of any the correspondence between points located on uniform areas of the images (e.g. regions with uniform colour or texture). This paper presents a new robust approach for establishing any correspondence between arbitrarily selected points in two widely-baselined views, based on the ASIFT (affine scale-invariant feature transform) method, image segmentation, and local homography. This method, denoted as ASIFT-SH, consists of the following steps: (i) determination of initial corresponding points, (ii) grouping of initial corresponding points into subsets, (iii) calculation of local homographies for a particular subset of points, and (iv) determination of any correspondence between arbitrary points from a particular cameras view by using a suitable local homography. The ASIFT-SH method, when compared to searching the area surrounding an epipolar line (EPI method), provides more accurate results, especially on surfaces with similar pixel intensities. The average error in our method it is in the order of a few pixels, whilst for the EPI method in the order of a few hundred pixels.

Computerized detection and recognition of follicles in ovarian ultrasound images: a review

Observing changes in females’ ovaries is essential in obstetrics and gynaecological imaging, e.g., genetic engineering and human reproduction. It is particularly important to monitor the dynamics of ovarian follicles’ growth, as only fully mature and grown follicles, i.e., the dominant follicles have a potential to ovulate at the end of a follicular phase. Gynaecologists follow this process in two dimensions, but recently three-dimensional (3-D) ultrasound examinations are coming to the fore. This paper surveys the existing computer methods for detection, recognition, and analyses of follicles in two-dimensional (2-D) and 3-D ovarian ultrasound recordings. Our study focuses on the efficiency, validation, and assessment of proposed follicle processing algorithms. The most important processing steps were identified in order to compare their performances. Higher ranking solutions are suggested for the so-called best algorithm for 2-D and 3-D ultrasound recordings of ovarian follicles. Finally, some guidelines for future research in this field are discussed, in particular for 3-D ultrasound volumes.

Soil-moisture estimation from TerraSAR-X data using neural networks

Early prediction of natural disasters like floods and landslides is essential for reasons of public safety. This can be attained by processing Synthetic-Aperture Radar (SAR) images and retrieving soil-moisture parameters. In this article, TerraSAR-X product images are investigated in combination with a water-cloud model based on the Shi semi-empirical model to determine the accuracy of soil-moisture parameter retrieval. SAR images were captured between January 2008 and September 2010 in the vicinity of the city Maribor, Slovenia, at different incidence angles. The water-cloud model provides acceptable estimated soil-moisture parameters at bare or scarcely vegetated soil areas. However, this model is too sensitive to speckle noise; therefore, a pre-processing step for speckle-noise reduction is carried out. Afterwards, self-organizing neural networks (SOM) are used to segment the areas at which the performance of this model is poor, and at the same time neural networks are also used for a more accurate approximation of model parameters’ values. Ground-truth is measured using the Pico64 sensor located on the field, simultaneously with capturing SAR images, in order to enable the comparison and validation of the obtained results. Experimental results show that the proposed method outperforms the water-cloud model accuracy over all incidence angles.

Improved prediction-based ovarian follicle detection from a sequence of ultrasound images

A new algorithm is presented for ovarian follicle recognition from a sequence of ultrasound images. The basic version of the prediction-based algorithm is upgraded by means of two improvements. The negative influence brought by the gross measurement errors is suppressed, and the locality of the treated process is considered. The basis for both improvements is the Kalman filter. The proposed algorithm is a combination of three mutually dependent Kalman filters: a global one whose parameters are then modified by two additional ones, firstly detecting the gross measurement errors and secondly, regarding the recognised contour of the object. The obtained results show that the follicles recognised using the final prediction algorithm are about 2% more compact and about 6% more accurate, on average, when compared to the values obtained using the basic prediction-based algorithm.

Assessing the influence of temperature variations on the geometrical properties of a low-cost calibrated camera system by using computer vision procedures

This paper estimates temperature influence on geometrical properties of both a single camera and a calibrated camera system, assuming low-cost CCD cameras. It does not cover the effect of temperature on the camera’s electronics. Firstly, the influence of temperature change on camera parameters was modelled and integrated into an existing analytical camera model. A modified camera model enables quantitative assessment regarding the influence of temperature variations for a single camera. Temperature variations also directly influence the accuracies of calibrated cameras. The inability to analytically determine the calibration method error magnitude, led us to experimentally estimate errors regarding calibrated cameras. Finally, the total error regarding calibrated cameras was derived by combining the numerical error of the calibration method with those errors originating from temperature variations. The results show that the influence of temperature variations decreases when increasing the distances of the observed objects from the cameras. On a typical building site, the temperature influence is reflected in the image as an error of less than one pixel.

Fast Segmentation of Ovarian Ultrasound Volumes Using Support Vector Machines and Sparse Learning Sets

Ovarian ultrasound imaging has recently drawn attention because of the improved ultrasound-based diagnostic methods and because of its application to in-vitro fertilisation and prediction of women’s fertility. Modern ultrasound devices enable frequent examinations and sophisticated built-in image processing options. However, precise detection of different ovarian structures, in particular follicles and their growth still need additional, mainly off-line processing with highly specialised algorithms. Manual annotation of a whole 3D ultrasound volume consisting of 100 and more slices, i.e. 2D ultrasound images, is a tedious task even when using handy, computer-assisted segmentation tools. Our paper reveals how an application of support vector machines (SVM) can ease the follicle detection by speeding up the learning and annotation processes at the same time. An iterative SVM approach is introduced using training on sparse learning sets only. The recognised follicles are compared to the referential expert readings and to the results obtained after learning on the entire annotated 3D ovarian volume.