My new scientific results, where my contribution was essential, will be summarized in three thesis points. All of them have been achieved after obtaining my PhD degree. Related chapters and relevant publications of the dissertation are listed at the end of each thesis point.
1. Optimal sampling and compression of objects
A possible way of object detection is to match predefined templates to the processed image.
In real time applications, the required time of the matching procedure is a key factor. The compact representation of the matched templates is also an important issue for their efficient storage in devices having limited resources. In this field, I have introduced methods for object sampling and compression based on theoretical results partly obtained by me; I have also verified their efficiency experimentally.
(i) By the extension on the theory of the centroidal Voronoi tessellation (CVT) [92, 93] I have introduced a new sampling procedure for the simplification of objects in both dis-crete and Euclidean domains. The sampling error is defined differently as in the classic CVT theory. Namely, it is calculated as the integral of distances between the object and sampled points within a dilated variant of the object. I have proven that this technique is optimal in object detection methods using the chamfer matching framework. I have shown the applicability of the approach for region-based representations of the objects besides their contour-based ones. Especially, I have introduced a weight function to concentrate the sampling of the object to its skeleton. The results have been integrated in the thermal video processing module of the system SHARE to detect human appear-ances. To summarize, I have shown the efficiency of the proposed sampling method by theoretically proving its optimal behavior and by verifying it experimentally. See Chapter 2. Related publications: in SCI1 journal: [1], other: [17–19].
(ii) I have composed a novel technique for the compression of digital curves having arbit-rary topology. Opposite to other approaches, the proposed method traces the whole curve, so it is capable of building it up from an alphabet of short line segments. I have given graph theoretical algorithms for the tracing of the curves providing efficient compression. Namely, a curve can be traced by decomposing it to Eulerian graphs or, to avoid decomposition, by adopting the solution of the Chinese Postman Problem, where the weights relate to the compression demands of the specific curve segments. I have introduced a procedure based on a new theoretical model to detect and improve the degenerated digital intersections turning up during the skeletonization of thicker curves. This model interprets the crossing segments as lanes and calculates the volume of deformation in terms of a function of the enclosed angle and widths of these lanes.
The curve compression technique has been integrated in the system SHARE to store its template database more efficiently. To summarize, I have shown the efficiency of a
1Science Citation Index.
novel curve compression method built upon theoretical derivations; I have performed experimental verification, as well. See Chapter 3. Related publications: in SCI journals:
[2], other: [20, 21].
2. Detection of objects having single or multiple appearances
A common image processing task is to reliably recognize and locate the appearances of specific objects. Besides the form of their appearances, like shape and color, we usually have some prior information about the possible number (single or multiple) of appearing objects of certain types. One of the possible approaches to increase the accuracy of object detection is to take as many principles as possible into account from the ones that have been developed for the given problem. In this field, I have introduced novel voting models for the spatial domain; I have proven their validity by theoretical tools and also experimentally.
(i) I have introduced fusion-based techniques for the detection of single objects that can be represented by single pixels. The output of different individual object detector al-gorithms are aggregated using simple or weighted majority voting-based models. I have made several efforts to improve the detection accuracy. I have composed a graph theoretical procedure for the simultaneous detection of more objects having single ap-pearances with incorporating a constraint regarding their relative spatial locations. I have given also a graph theoretical method that selects the maximum weighted cliques to handle the problem, when the detectors can have more candidates for the location of the object. To take advantage of all the information from the outputs of the individ-ual detectors, I have introduced aggregation models based on axiomatic and Bayesian models. The dependencies between the detectors have been discovered by statistical methods based on the individual variances and pairwise covariances. The dependency issue has been addressed by assigning weights to the outputs, where the weights have been determined via solving a minimization problem for the variance of the final can-didate. These approaches have been implemented in the system DRSCREEN to detect the optic disc and macula in retinal images. To summarize, I have experimentally vali-dated the efficiency of a new fusion-based model partly having theoretical foundations.
See Chapter 4. Related publications: in SCI journals: [3, 4], other: [22, 23].
(ii) I have generalized the classic majority and weighted majority voting models to the spa-tial domain, where the member algorithms vote in terms of pixels for the location of an object. For this purpose, I have introduced a probability term, which can be specifically defined by a geometric constraint according to the shape of the object. I have described the detection accuracy behavior of the new model as a function of the individual ac-curacies of the member algorithms and the dependencies among them. Namely, I have proven new theoretical results for the minimal and maximal independence of a set of random variables using linear programming. I have determined the change of the accu-racy of the ensemble in that practically essential case, when a new member algorithm is added to the system. For disc-like objects, I have derived the behavior of the prob-ability term of the generalized voting model with improving former theoretical results on the diameter of sets of point randomly dropped on a disc. Namely, I have shown that the probability of having the wrong outputs within a disc drops exponentially with the number of detectors in the independent case. I have extended the classic weighted majority voting model to a novel spatial voting one; the appropriate weights to detect disc-like object have been determined by the generalization of the recommendations for the classic case. To discover better the dependencies within the spatial voting model,
I have generalized the classic diversity measures and given a technique to compose ef-ficient ensembles using them. I have demonstrated the applicability of the new models for the detection of the optic disc; the results have been implemented in the system DRSCREEN. To summarize, I have proven the suitability of a novel spatial voting model for the detection of objects that can be represented by single pixels. See Chapter 5. Related publications: in SCI journal: [5], other: [24–26].
(iii) I have introduced a novel fusion-based technique for the detection of objects having multiple appearances in the image. The method increases the accuracy of the ensemble by making its members more diverse via composing them as pairs of preprocessing algo-rithms and candidate extractors. To build up an ensemble of such pairs with maximum accuracy, I have composed a stochastic search-based algorithm and tested it exhaus-tively for energy functions regularly considered to measure the error of detection. I have verified the efficiency of the approach for the detection of microaneurysms in reti-nal images. To raise further the detection accuracy of the ensemble, I have worked out a context-aware method which takes the appearance features and spatial location of the objects also into account. The results have been implemented in the system DRSCREEN. To summarize, I have shown that the proposed method to make ensem-bles more diverse is efficient for object detection purposes. See Chapter 6. Related publications: in SCI journals: [6–8], other: [27–32].
3. Ensemble-based systems in decision making
Just in the object detection task, it is worth taking several principles and opinions into consideration to support decisions. The related classic approach is to extract features and to make a decision according to them using a machine learning-based classifier. Besides the high quality features, for the final decision it may be also useful to combine classifiers considering different principles. In this field, I have introduced a new method for the fusion of classifiers; I have validated the approach experimentally within a complex application.
(i) I have introduced a new ensemble-based decision support model, which consists of classic classification methods. The fusion of the classifiers is achieved via axiomatic, majority and weighted majority voting-based models. I have given a method to select the appropriate classifiers and fusion rules to compose ensembles having highest de-tection accuracies regarding different energy functions. I have verified the efficiency of the approach for both binary and multiclass classification problems according to the requirements of an application field. Namely, the new decision support approach have been implemented in the system DRSCREEN dedicated to the automatic screening of diabetic retinopathy based on retinal images. To summarize, I have shown that the new model considering the fusion of classifiers is efficient for decision support purposes. See Chapter 7. Related publications: in SCI journal: [9], other: [33, 34].
(ii) I have achieved several results corresponding to the improvement of some components and features of the ensemble-based system discussed in the previous point. Beyond the results presented in the thesis points so far, I have created new methods to exclude obviously diseased or low quality images from further processing based on local inhomo-geneity descriptors. I have given a new technique applying mathematical morphology for the detection of the macula, and other ones for the extraction of the retinal vessels using hidden Markov random fields and the detection of microaneurysms through the analysis of intensity profiles. As for the latter field, I have introduced several new fea-tures extracted from the functional representations of the intensity profiles for machine
learning-based frameworks. Using a similar principle as for the detection of microa-neurysms, I have introduced an ensemble-based system for the detection of region-like objects with verifying the efficiency of this approach for the detection of exudates, which lesions are also specific to diabetic retinopathy. These results have been implemented in the corresponding components of the system DRSCREEN. To summarize, I have shown how the accuracy of a complex decision support system can be raised by improving the efficiency of its components. See Chapter 7. Related publications: in SCI journal: [10], other: [35–38].
I have achieved several research results related to the ones presented in the dissertation. However, they are not discussed in details and are also excluded from the thesis points partly because of space reasons and to make the dissertation compact. See Chapter 8. Related publications: in SCI journals: [11–16],other: [39–55].