In traditional classification problems, all classes in the test set are assumed to also occur in the training set, also referred to as the closed-set assumption. However, in practice, new classes may occur in the test set, which reduces the performance of machine learning models trained under the closed-set assumption. Machine learning models should be able to accurately classify instances of classes known during training while concurrently recognizing instances of previously unseen classes (also called the open set assumption). This open set assumption is motivated by real world applications of classifiers wherein its improbable that sufficient data can be collected a priori on all possible classes to reliably train for them. For example, motivated by the DARPA WASH project at WPI, a disease classifier trained on data collected prior to the outbreak of COVID-19 might erroneously diagnose patients with the flu rather than the novel coronavirus. State-of-the-art open set methods based on the Extreme Value Theory (EVT) fail to adequately model class distributions with unequal variances. We propose the Variational Open-Set Recognition (VOSR) model that leverages all class-belongingness probabilities to reject unknown instances. To realize the VOSR model, we design a novel Multi-Modal Variational Autoencoder (MMVAE) that learns well-separated Gaussian Mixture distributions with equal variances in its latent representation. During training, VOSR maps instances of known classes to high-probability regions of class-specific components. By enforcing a large distance between these latent components during training, VOSR then assumes unknown data lies in the low-probability space between components and uses a multivariate form of Extreme Value Theory to reject unknown instances. Our VOSR framework outperforms state-of-the-art open set classification methods with a 15% F1 score increase on a variety of benchmark datasets.

No description available.

Thesis (MSc (Geography and Environmental Studies))--University of Stellenbosch, 2010. / ENGLISH ABSTRACT: Supervised classifiers are the most popular approach for image classification due to their high accuracies, ease of use and strong theoretical grounding. Their primary disadvantage is the high level of user input required during the creation of the data needed to train the classifier. One alternative to supervised classification is an expert-system rule-based approach where expert knowledge is used to create a set of rules which can be applied to multiple images. This research compared supervised and expert-system rule-based approaches for forest mapping. For this purpose two SPOT 5 images were acquired and atmospherically corrected. Field visits, aerial photography, high resolution imagery and expert forestry knowledge were used for the compilation of the training data and the development of a rule-set. Both approaches were evaluated in an object-orientated environment. It was found that the accuracy of the resulting maps was equivalent, with both techniques returning an overall classification accuracy of 90%. This suggests that cost-effectiveness is the decisive factor for determining which method is superior. Although the development of the rule-set was time-consuming and challenging, it did not require any training data. In contrast, the supervised approach required a large number of training areas for each image classified, which was time-consuming and costly. Significantly more training areas will be required when the technique is applied to large areas, especially when multiple images are used. It was concluded that the rule-set is more cost-effective when applied at regional scale, but it is not viable for mapping small areas. / AFRIKAANSE OPSOMMING: Gerigte klassifiseerders is die gewildste benadering tot beeldklassifikasie as gevolg van hulle hoë graad van akkuraatheid, maklike aanwending en kragtige teoretiese fundering. Die primere nadeel van gerigte klassifikasie is die hoë vlak van gebruikersinsette wat benodig word tydens die skepping van opleidingsdata. 'n Alternatief vir gerigte klassifikasie is 'n deskundige stelsel waarin ‘n reëlgebaseerde benadering gevolg word om deskundige kennis aan te wend vir die opstel van 'n stel reëls wat op meervoudige beelde toegepas kan word. Hierdie navorsing het gerigte en deskundige stelsel benaderings toegepas vir bosboukartering om die twee benaderings met mekaar te vergelyk. Vir dié doel is twee SPOT 5 beelde verkry en atmosferies gekorrigeer. Veldbesoeke, lugfotografie, hoë-resolusie beelde en deskundige bosboukennis is aangewend om opleidingsdata saam te stel en die stel reëls te ontwikkel. Beide benaderings is in 'n objekgeoriënteerde omgewing beoordeel. Die akkuraatheidsvlakke van die resulterende kaarte was ewe hoog vir beide tegnieke met 'n algehele klassifikasie-akkuraatheid van 90%. Dit wil dus voorkom asof koste-effektiwiteit eerder as akkuraatheid die deurslaggewende faktor is om te bepaal watter metode die beste is. Alhoewel die ontwikkeling van die stel reëls tydrowend en uitdagend was, het dit geen opleidingsdata vereis nie. In teenstelling hiermee is 'n groot aantal opleidingsgebiede geskep vir elke beeld wat met gerigte klassifikasie verwerk is – 'n tydrowende en duur opsie. Dit is duidelik dat meer opleidingsgebiede benodig sal word wanneer die tegniek op groot gebiede toegepas word, veral omdat meervoudige beelde gebruik sal word. Gevolglik sal die stel reëls meer kosteeffektief wees wanneer dit op streekskaal toegepas word. ‘n Deskundige stelsel benadering is egter nie lewensvatbaar vir die kartering van klein gebiede nie.

Given the strong association between aberrant nuclear morphology and tumor progression, changes in nuclear structure have remained the gold standard for cancer diagnosis for over 150 years. Recently, the rapid development of imaging hardware and computation power creates the opportunity for automated computer-aided diagnosis (CAD). Developing a robust and reliable pattern recognition pipeline is a pressing need to mine and analyze tons of nuclei data being captured. Among the rich studies on pattern recognition problems in pathology, automated nuclei detection, segmentation and cancer detection are the recurring tasks due to the importance and challenges of nuclei analysis. In this thesis, we propose and investigate the state-of-art methods in the CAD modules for maximizing the overall amount of information from images for decision making. We focus on nuclei segmentation and patient cancer detection in the nuclei image analysis pipeline. As the first step in nuclei analysis, we develop an unsupervised nuclei detection and segmentation approach for pathology images. Different from many supervised segmentation methods whose performances rely on the quality and quantity of training samples, the proposed method is able to automatically search for the nucleus contour by solving the shortest path problem with little user effort. We consider the cancer detection task as a set classification problem and propose a highly discriminative predictive model in the sense that it not only optimizes the classifier decision boundary but also transfers discriminative information to set representation learning. The innovation of the model is the integration of set representation learning and classifier training into one objective function for boosting the cancer detection performance. Experimental results showed that the new model provides significant improvements compared with state-of-art methods in the diagnostic challenges. In addition, we showed that the predictive model enables visual interpretation of discriminative nuclear characteristics representing the whole nuclei set. We believe the proposed model is quite general and provide experimental validations in several extended pattern recognition problems.

No description available.

The increase of mobile smartphones continues to grow and with it the demand for automation and use of the most offered aspects of the phone, whether in medicine (health care and surveillance) or in user applications (automatic recognition of position, etc.). As part of this work has been created the designs and implementation of the system for the recognition of human activity on the basis of data processing from sensors of smartphones, along with the determination of the optimal parameters, recovery success rate and comparison of individual evaluation. Other benefits include a draft format and displaying numerous training set consisting of real contributions and their manual evaluation. In addition to the main benefits, the software tool was created to allow the validation of the elements of the training set and acquisition of features from this set and software, that is able with the help of deep learning to train models and then test them.