Search

Effective training of filters in Convolutional Neural Networks (CNN) ensures their success. In order to achieve good classification results in CNNs, filters must be carefully initialized, trained and fine-tuned. We propose an unsupervised method that allows the discovery of filters from the given dataset in a single epoch without specifying the number of filters hyper-parameter in convolutional layers. Our proposed method gradually builds the convolutional layers by a discovery routine that extracts a number of features that adequately represent the complexity of the input domain. The discovered filters represent the patterns in the domain, so they do not require any initialization method or backpropagation training for fine tuning purposes. Our method achieves 99.03% accuracy on MNIST dataset without applying any data augmentation techniques. Publisher's Version

National mapping agencies (NMAs) have to acquire nation-wide Digital Terrain Models on a regular basis as part of their obligations to provide up-to-date data. Point clouds from Airborne Laser Scanning (ALS) are an important data source for this task; recently, NMAs also started deriving Dense Image Matching (DIM) point clouds from aerial images. As a result, NMAs have both point cloud data sources available, which they can exploit for their purposes. In this study, we investigate the potential of transfer learning from ALS to DIM data, so the time consuming step of data labelling can be reduced. Due to their specific individual measurement techniques, both point clouds have various distinct properties such as RGB or intensity values, which are often exploited for classification of either ALS or DIM point clouds. However, those features also hinder transfer learning between these two point cloud types, since they do not exist in the other point cloud type. As the mere 3D point is available in both point cloud types, we focus on transfer learning from an ALS to a DIM point cloud using exclusively the point coordinates. We are tackling the issue of different point densities by rasterizing the point cloud into a 2D grid and take important height features as input for classification. We train an encoder-decoder convolutional neural network with labelled ALS data as a baseline and then fine-tune this baseline with an increasing amount of labelled DIM data. We also train the same network exclusively on all available DIM data as reference to compare our results. We show that only 10% of labelled DIM data increase the classification results notably, which is especially relevant for practical applications.

Tactile data processing and analysis is still essentially an open challenge. In this framework, we demonstrate a method to achieve touch modality classification using pre-trained convolutional neural networks (CNNs). The 3D tensorial tactile data generated by real human interactions on an electronic skin (E-Skin) are transformed into 2D images. Using a transfer learning approach formalized through a CNN, we address the challenging task of the recognition of the object that was touched by the E-Skin. The feasibility and efficiency of the proposed method are proven using a real tactile dataset outperforming classification results obtained with the same dataset in the literature.

We propose a novel hierarchical approach for the classification of airborne 3D lidar points. Spatial and semantic context is incorporated via a two-layer Conditional Random Field (CRF). The first layer operates on a point level and utilises higher order cliques. Segments are generated from the labelling obtained in this way. They are the entities of the second layer, which incorporates larger scale context. The classification result of the segments is introduced as an energy term for the next iteration of the point-based layer. This framework iterates and mutually propagates context to improve the classification results. Potentially wrong decisions can be revised at later stages. The output is a labelled point cloud as well as segments roughly corresponding to object instances. Moreover, we present two new contextual features for the segment classification: the distance and the orientation of a segment with respect to the closest road. It is shown that the classification benefits from these features. In our experiments the hierarchical framework improve the overall accuracies by 2.3% on a point-based level and by 3.0% on a segment-based level, respectively, compared to a purely point-based classification.

Spray drift is one of the main pollution sources identified when pesticides are sprayed on crops. In this work, in order to simplify the evaluation of hollow-cone nozzles according to their drift potential reduction, several models commonly used were tested by three indirect methods: phase Doppler particle analyser (PDPA) and two different wind tunnels. The main aim of this study is then to classify for the first time these hollow-cone nozzle models all of them used in tree crop spraying (3D crops). A comparison between these indirect methods to assess their suitability and to provide guidelines for a spray drift classification of hollow-cone nozzles was carried out. The results show that, in general terms, all methods allow hollow-cone nozzle classifications according to their drift potential reduction (DPR) with a similar trend. Among all the parameters determined with the PDPA, the V100 parameter performed best in differentiating the tested nozzles among drift reduction classes. In the wind tunnel, similar values were obtained for both sedimenting and airborne drift depositions. The V100 parameter displayed a high correlation (up to R2 = 0.948) with the drift potential tested with the wind tunnel. It is concluded that in general, the evaluated indirect methods provide equivalent classification results. Additional studies with a greater variety of nozzle types are required to achieve a proposal of harmonized methodology for testing hollow-cone nozzles. This work was partly funded by the Secretaria d'Universitats i Recerca del Departament d'Empresa i Coneixement de la Generalitat de Catalunya, the Spanish Ministry of Economy and Competitiveness and the European Regional Development Fund (ERDF) under Grants 2017 SGR 646, AGL2007-66093-C04-03, AGL2010-22304-04-C03-03, and AGL2013-48297-C2-2-R. The authors also wish to thank Mr. Antonio Checa (Randex Iberica, S.L.) for giving us free Albuz nozzles for the spray tests. Universitat de Lleida is also thanked for Mr. X. Torrent's pre-doctoral fellowship.

In this article, we study various analytic aspects of interpolating sesqui-harmonic maps between Riemannian manifolds where we mostly focus on the case of a spherical target. The latter are critical points of an energy functional that interpolates between the functionals for harmonic and biharmonic maps. In the case of a spherical target, we will derive a conservation law and use it to show the smoothness of weak solutions. Moreover, we will obtain several classification results for interpolating sesqui-harmonic maps.

In the context of mobile Health (mHealth) applications, data are prone to several sources of contamination which would lead to false interpretation and misleading classification results. In this paper, a robust deep learning approach with low rank model is proposed to classify mHealth vital signs. Further-more, we propose using the Schatten-p norm instead of the classic nuclear norm since it has shown better recovery performance for several applications. We conduct a comprehensive study where we compare our method to the state-of-art methods and evaluate its performance with respect to the key system parameters. Our findings show indeed that combining deep network with dictionary learning model is effective for vital signs classification even in presence of 50% corruption with 8% improvement over the closest performance. 2018 IEEE. Qatar Foundation;Qatar National Research Fund Scopus

In this paper a suitable methodology for the improvement of the reliability of results in classification systems based on 3D images is proposed. More in detail, it is based on the knowledge of the uncertainty of the features constituting the 3D image (obtained processing a pair of two 2D stereoscopic images) and on a suitable statistical approach providing a confidence level to the classification result. These pieces of information are then managed in order to improve the classification performance in terms of correct classification and false reject percentages. The experimental results, obtained applying the methodology on an Active Appearance Models algorithm for feature detection and triangulating the 3D features, show that, compared with a basic approach (which generally does not take into account the uncertainty on 3D features), the proposed methodology allows to significantly improve the classification performance even in scenarios characterized by a high uncertainty.

Date of Conference: 28 Aug.-2 Sept. 2017 Conference name: 25th European Signal Processing Conference (EUSIPCO), 2017 We introduce a time-frequency scattering method using hyperbolic tangent function for vessel sound classification. The sound data is wavelet transformed using a two channel filter-bank and filter-bank outputs are scattered using tanh function. A feature vector similar to mel-scale cepstrum is obtained after a wavelet packed transform-like structure approximating the mel-frequency scale. Feature vectors of vessel sounds are classified using a support vector machine (SVM). Experimental results are presented and the new feature extraction method produces better classification results than the ordinary Mel-Frequency Cepstral Coefficients (MFCC) vectors. © EURASIP 2017.

Updating topographic geospatial databases is often performed based on current remotely sensed images. To automatically extract the object information (labels) from the images, supervised classifiers are being employed. Decisions to be taken in this process concern the definition of the classes which should be recognised, the features to describe each class and the training data necessary in the learning part of classification. With a view to large scale topographic databases for fast developing urban areas in the Kingdom of Saudi Arabia we conducted a case study, which investigated the following two questions: (a) which set of features is best suitable for the classification?; (b) what is the added value of height information, e.g. derived from stereo imagery? Using stereoscopic GeoEye and Ikonos satellite data we investigate these two questions based on our research on label tolerant classification using logistic regression and partly incorrect training data. We show that in between five and ten features can be recommended to obtain a stable solution, that height information consistently yields an improved overall classification accuracy of about 5%, and that label noise can be successfully modelled and thus only marginally influences the classification results.

Brain activity during perception and recognition of faces have been studied by researchers with the purpose to develop brain-computer interfaces and to study neurological disorders. In this paper, we analyzed evoked potentials as neurophysiological indicators and developed a model based on signal processing and machine learning techniques to find descriptive patterns that allow the differentiation of familiar and unfamiliar faces. We considered wave components such as P1, N170, N250, P300, and N400 to describe the events. Morphological analysis and wavelet transform were used for the feature extraction stage, and support vector machines and binomial logistic regression were evaluated for the classification stage. The best classification results were obtained with the morphological characteristics, where the highest classification accuracy was 80% on average. © 2018 IEEE. Institute of Electrical and Electronics Engineers Colombia Section;Institute of Electrical and Electronics Engineers Consejo Andino

In this thesis, we have studied the problem of detecting automated Twitter accounts related to the Ukraine conflict using supervised learning. A striking problem with the collected data set is that it was initially lacking a ground truth. Traditionally, supervised learning approaches rely on manual annotation of training sets, but it incurs tedious work and becomes expensive for large and constantly changing collections. We present a novel approach to synthetically generate large amounts of labeled Twitter accounts for detection of automation using a rule-based classifier. It significantly reduces the effort and resources needed and speeds up the process of adapting classifiers to changes in the Twitter-domain. The classifiers were evaluated on a manually annotated test set of 1,000 Twitter accounts. The results show that rule-based classifier by itself achieves a precision of 94.6% and a recall of 52.9%. Furthermore, the results showed that classifiers based on supervised learning could learn from the synthetically generated labels. At best, the these machine learning based classifiers achieved a slightly lower precision of 94.1% compared to the rule-based classifier, but at a significantly better recall of 93.9% Detta exjobb har undersökt problemet att detektera automatiserade Twitter-konton relaterade till Ukraina-konflikten genom att använda övervakade maskininlärningsmetoder. Ett slående problem med den insamlade datamängden var avsaknaden av träningsexempel. I övervakad maskininlärning brukar man traditionellt manuellt märka upp en träningsmängd. Detta medför dock långtråkigt arbete samt att det blir dyrt förstora och ständigt föränderliga datamängder. Vi presenterar en ny metod för att syntetiskt generera uppmärkt Twitter-data (klassifieringsetiketter) för detektering av automatiserade konton med en regel-baseradeklassificerare. Metoden medför en signifikant minskning av resurser och anstränging samt snabbar upp processen att anpassa klassificerare till förändringar i Twitter-domänen. En utvärdering av klassificerare utfördes på en manuellt uppmärkt testmängd bestående av 1,000 Twitter-konton. Resultaten visar att den regelbaserade klassificeraren på egen hand uppnår en precision på 94.6% och en recall på 52.9%. Vidare påvisar resultaten att klassificerare baserat på övervakad maskininlärning kunde lära sig från syntetiskt uppmärkt data. I bästa fall uppnår dessa maskininlärningsbaserade klassificerare en något lägre precision på 94.1%, jämfört med den regelbaserade klassificeraren, men med en betydligt bättre recall på 93.9%.

Detection of mitotic cells in histology images is an important but challenging process due to the resemblance of mitotic cells with other non-mitotic cells and also due to the different appearance of mitotic cells undergoing different phases of the division process. In this paper, we present an algorithm for classification of mitotic cells into its four different phases using eigenphase nuclei images - nuclear exemplars obtained separately from the eigen-decomposition of training nuclei images belonging to each of the four mitotic phases. To the best of our knowledge, ours is the first method to identify mitotic phases in cancer histology images. It is quite likely that the classification results may be negatively affected if the dataset used for training purposes does not contain sufficient number of samples for a positive class. To overcome this class imbalance problem, we present a novel method for oversampling the minority class. The proposed method generates synthetic images for training purposes by perturbing the representation of training samples belonging to the minority class in the eigenphase domain. We show that this strategy works effectively for pairwise classification of the mitotic cells - increasing the classification performance by as much as 24%. 2016 IEEE. Scopus

The aim of this study was to investigate the usefulness of multilevel binary and gray scale morphological analysis in the assessment of atherosclerotic carotid plaques. Ultrasound images were recorded from 137 asymptomatic and 137 symptomatic plaques (Stroke, Transient Ischaemic Attack (TIA), Amaurosis Fugax (AF)). We carefully develop the clinical motivation behind our approach. We do this by relating the proposed L-images, M-images and H-images in terms of the clinically established hypoechoic, isoechoic and hyperechoic classification. Normalized pattern spectra were computed for both a structural, multilevel binary morphological model, and a direct gray scale morphology model. From the plots of the average pattern spectra, it is clear that we have significant differences between the symptomatic and asymptomatic spectra. Here, we note that the morphological measurements appear to be in agreement with the clinical assertion that symptomatic plaques tend to have large lipid cores while the asymptomatic plaques tend to have small lipid cores. The derived pattern spectra were used as classification features with two different classifiers, the Probabilistic Neural Network (PNN) and the Support Vector Machine (SVM). Both classifiers were used for classifying the pattern spectra into either a symptomatic or an asymptomatic class. The highest percentage of correct classifications score was 73.7% for multilevel binary morphological image analysis and 66.8% for gray scale morphological analysis. Both were achieved using the SVM classifier. Among all features, the L-image pattern spectra, that also measure the distributions of the lipid core components (and some non-lipid components) gave the best classification results. © 2007 Springer Science+Business Media, LLC. 30 1 3 23 Cited By :49

In this paper a suitable methodology for the improvement of the reliability of results in classification systems based on 3D images is proposed. More in detail, it is based on the knowledge of the uncertainty of the features constituting the 3D image (obtained processing a pair of two 2D stereoscopic images) and on a suitable statistical approach providing a confidence level to the classification result. These pieces of information are then managed in order to improve the classification performance in terms of correct classification and missed classification percentages. The experimental results, obtained applying the methodology on an Active Appearance Models algorithm, a popular method for face recognition based on 3D features, show that, compared with a traditional approach (which generally does not take into account the uncertainty on 3D features), the proposed methodology allows to significantly improve the classification performance even in scenarios characterized by a high uncertainty.

This paper proposes a two-layer Conditional Random Field model for simultaneous classification of land cover and land use. Both classification tasks are integrated into a unified graphical model, which is reasonable due to the fact that land cover and land use exhibit strong contextual dependencies. In the CRF, we distinguish a land cover layer and a land use layer. Both layers differ with respect to the entities corresponding to the nodes and the classes to be distinguished. In the land cover layer, the nodes correspond to superpixels extracted from the image data, whereas in the land use layer the nodes correspond to objects of a geospatial land use database. Statistical dependencies between land cover and land use are explicitly modelled as pair-wise potentials. Thus, we obtain a consistent model, where the relations between land cover and land use are learned from representative training data. The approach is designed for input data based on aerial images. Experiments are performed on an urban test site. The experiments show the feasibility of the combination of both classification tasks into one overall approach and investigate the influence of the size of the superpixels on the classification result.

Date of Conference: 24-26 April 2013 In this work, we study the task of recognizing human actions from noisy videos and effects of noise to recognition performance and propose a possible solution. Datasets available in computer vision literature are relatively small and could include noise due to labeling source. For new and relatively big datasets, noise amount would possible increase and the performance of traditional instance based learning methods is likely to decrease. In this work, we propose a multiple instance learning-based solution in case of an increase in noise. For this purpose, each video is represented with spatio-temporal features, then bag-of-words method is applied. Then, using support vector machines (SVM), both instance-based learning and multiple instance learning classifiers are constructed and compared. The classification results show that multiple instance learning classifiers has better performance than instance based learning counterparts on noisy videos. © 2013 IEEE.

Dandanes se vsi trudimo, da naše poslovne ideje čim uspešnejše predstavimo potencialnim kupcem. Eden od pogosto rabljenih vzvodov je svetovni internet. Vendar zaradi velike konkurence ni več dovolj zgolj odlična internetna stran z vsemi informacijami, ki jih potencialni kupci potrebujejo, če ne najdejo do nje. Google in njemu podobni iskalniki krojijo poslovno uspešnost mnogim podjetjem širom sveta. Torej lahko govorimo v pravem boju za prikaz na prvi strani vseh zadetkov. Ravno za to je potrebno izkoristiti vse načine, katere nam omogoča napredek v tehnologiji. Eden od teh načinov so socialni mediji, katere je potrebno sistematično vpeljati v našo stran. Govorimo o socialnih medijih kot so: Facebook, Youtube, Flicker, Twitter in še bi lahko naštevali. Ti se nenehno razvijajo in rastejo. Poslovanje pa praviloma raste vzporedno z njimi, če jim le namenimo dovolj pozornosti. Pri sami optimizaciji s socialnimi mediji pa se je potrebno držati nekateri pravil, da dosežemo želeni učinek. Predvsem pa je potreben čas in veliko potrpežljivosti. Ob upoštevanju vseh smernic, kako uspešno vplesti uporabnike v samo zgodbo, uspeh prav gotovo ne bo izostal! Nowadays we are all trying to successfully represent our business ideas to our potential customers. One of the most common used instruments is worldwide web. Because of a great deal of competition it is nearly not enough to just have a good internet site with all the useful information for the potential customer, if they do not know how to direct themselves to it and find it. Google and similar search engines mark the path and make a difference between successful and less successful companies across the world. We may call it a web online war which takes place on the internet where the sites struggle to be able to reach the first positions results through the online search. Because of this situation we have to use all means necessary that technology offers to us today for achieving a better view of our web sites among others. A good way to achieve this are social Medias like: Facebook, Youtube, Flickr, Twitter, etc. These web communities are expanding, growing and evolving as we speak. Our businesses may grow with them if we just invest our attention in them. The optimization with social media needs a lot of time and patience investment but we follow all of the guidelines the business success will not fall behind.

Carotid plaques have been associated with ipsilateral neurological symptoms. High-resolution ultrasound can provide information not only on the degree of carotid artery stenosis but also on the characteristics of the arterial wall including the size and consistency of atherosclerotic plaques. The aim of this study is to determine whether the addition of ultrasonic plaque texture features to clinical features in patients with asymptomatic internal carotid artery stenosis (ACS) improves the ability to identify plaques that will produce stroke. 1121 patients with ACS have been scanned with ultrasound and followed for a mean of 4 years. It is shown that the combination of texture features based on second-order statistics spatial gray level dependence matrices (SGLDM) and clinical factors improves stroke prediction (by correctly predicting 89 out of the 108 cases that were symptomatic). Here, the best classification results of $77 \pm 1.8\%$ were obtained from the use of the SGLDM texture features with support vector machine classifiers. The combination of morphological features with clinical features gave slightly worse classification results of $76 \pm 2.6\%$. These findings need to be further validated in additional prospective studies. © 2012 IEEE. 16 5 966 973 Cited By :24

This study introduces the use of multiscale amplitude modulation-frequency modulation (AM-FM) texture analysis of multiple sclerosis (MS) using magnetic resonance (MR) images from brain. Clinically, there is interest in identifying potential associations between lesion texture and disease progression, and in relating texture features with relevant clinical indexes, such as the expanded disability status scale (EDSS). This longitudinal study explores the application of 2-D AM-FM analysis of brain white matter MS lesions to quantify and monitor disease load. To this end, MS lesions and normal-appearing white matter (NAWM) from MS patients, as well as normal white matter (NWM) from healthy volunteers, were segmented on transverse T2-weighted images obtained from serial brain MR imaging (MRI) scans (0 and 6-12 months). The instantaneous amplitude (IA), the magnitude of the instantaneous frequency (IF), and the IF angle were extracted from each segmented region at different scales. The findings suggest that AM-FM characteristics succeed in differentiating 1) between NWM and lesions 2) between NAWM and lesions and 3) between NWM and NAWM. A support vector machine (SVM) classifier succeeded in differentiating between patients that, two years after the initial MRI scan, acquired an EDSS ≤ 2 from those with EDSS > 2 (correct classification rate = 86%). The best classification results were obtained from including the combination of the low-scale IA and IF magnitude with the medium-scale IA. The AM-FM features provide complementary information to classical texture analysis features like the gray-scale median, contrast, and coarseness. The findings of this study provide evidence that AM-FM features may have a potential role as surrogate markers of lesion load in MS. © 2006 IEEE. 15 1 119 129 Cited By :31

We present the use of multiscale Amplitude Modulation Frequency Modulation (AM-FM) methods for analyzing brain white matter lesions that are associated with disease progression. We analyze lesions and normal appearing white matter (NAWM) longitudinally (0 and 6 months) and also for progression of disease. We use the expanded disability status scale (EDSS) to assess disease progression. The findings suggest that the high-frequency scale instantaneous amplitude can be used to differentiate between lesions associated with early and advanced disease stages. The classification results using the IF information and support vector machines produced a maximum sensitivity of 0.86, specificity of 0.76 and a maximum correct classification of 0.71. © 2010 IEEE. 61 64 Sponsors: The Institute of Electrical and Electronics Engineers IEEE Computer Society Conference code: 81155 Cited By :1