Auflistung nach Schlagwort "deep learning"
1 - 10 von 31
Treffer pro Seite
Sortieroptionen
- KonferenzbeitragActive-learning-driven deep interactive segmentation for cost-effective labeling of crop-weed image data(43. GIL-Jahrestagung, Resiliente Agri-Food-Systeme, 2023) Sikouonmeu, Freddy; Atzmueller, MartinActive learning has shown its reliability in (semi-)supervised machine learning tasks to reduce the labeling cost for large datasets in various areas. However, in the agricultural field, despite past attempts to reduce the labeling cost and the burden on the labeler in acquiring image labels, the load during the acquisition of pixel-level labels for semantic image segmentation tasks remains high. Typically, the respective pixel-level masks are acquired manually by drawing polygons over irregular and complex-shaped object boundaries. In contrast, this paper proposes a method leveraging the power of a click-based deep interactive segmentation model (DISEG) in an active learning approach to harvest high-quality image segmentation labels at a low cost for training a real-time task model by only clicking on the objects’ fore- and background surfaces. Our first experimental results indicate that with an average of 3 clicks per image object and using only 3% of the unlabeled dataset, we can acquire pixel-level labels with good quality at low cost.
- KonferenzbeitragAdversarial Attacks on Graph Neural Networks(INFORMATIK 2019: 50 Jahre Gesellschaft für Informatik – Informatik für Gesellschaft, 2019) Zügner, Daniel; Akbarnejad, Amir; Günnemann, Stephan
- TextdokumentApplying a deep learning-based approach for scaling vegetation dynamics to predict changing forest regimes under future climate and fire scenarios(INFORMATIK 2020, 2021) Rammer, Werner; Seidl, RupertThe ability to anticipate future changes in terrestrial ecosystems is key for their management. New tools are required that bridge the gap between a high level of process understanding at fine spatial grain, and the increasing relevance for management at larger extents. Such a tool is SVD (Scaling Vegetation Dynamics), a scaling framework that specifically uses deep learning to learn the behavior of detailed vegetation models in response to different environmental factors. This trained deep neural network (DNN) is then applied within the framework on large spatial scales. In addition, SVD includes also explicitly modelled processes such as fire disturbances. Here we use the framework to simulate forest regime change in the 3 Mio. ha landscape of the Greater Yellowstone Ecosystem. We used four climate change scenarios and pre-defined fire events from statistical modelling, and analyzed whether prevailing forest types are able to regenerate after fire. Our results show that up to 60% of the area may undergo regime change until the end of the 21st century.
- ZeitschriftenartikelBest low-cost methods for real-time detection of the eye and gaze tracking(i-com: Vol. 23, No. 1, 2024) Khaleel, Amal Hameed; Abbas, Thekra H.; Ibrahim, Abdul-Wahab SamiThe study of gaze tracking is a significant research area in computer vision. It focuses on real-world applications and the interface between humans and computers. Recently, new eye-tracking applications have boosted the need for low-cost methods. The eye region is a crucial aspect of tracking the direction of the gaze. In this paper, several new methods have been proposed for eye-tracking by using methods to determine the eye area as well as find the direction of gaze. Unmodified webcams can be used for eye-tracking without the need for specialized equipment or software. Two methods for determining the eye region were used: facial landmarks or the Haar cascade technique. Moreover, the direct method, based on the convolutional neural network model, and the engineering method, based on distances determining the iris region, were used to determine the eye’s direction. The paper uses two engineering techniques: drawing perpendicular lines on the iris region to identify the gaze direction junction point and dividing the eye region into five regions, with the blackest region representing the gaze direction. The proposed network model has proven effective in determining the eye’s gaze direction within limited mobility, while engineering methods improve their effectiveness in wide mobility.
- TextdokumentBidirectional Transformer Language Models for Smart Autocompletion of Source Code(INFORMATIK 2020, 2021) Binder, Felix; Villmow, Johannes; Ulges, AdrianThis paper investigates the use of transformer networks – which have recently become ubiquitous in natural language processing – for smart autocompletion on source code. Our model JavaBERT is based on a RoBERTa network, which we pretrain on 250 million lines of code and then adapt for method ranking, i.e. ranking an object's methods based on the code context. We suggest two alternative approaches, namely unsupervised probabilistic reasoning and supervised fine-tuning. The supervised variant proves more accurate, with a top-3 accuracy of up to 98%. We also show that the model – though trained on method calls' full contexts – is quite robust with respect to reducing context.
- KonferenzbeitragCan point-cloud based neural networks learn fingerprint variability?(BIOSIG 2022, 2022) Dominik Söllinger, Robert JöchlSubject- and environmental-specific variations affect the fingerprint recognition process. Quality metrics are capable of detecting and rating severe degradations. However, measuring natural variability, occurring during different fingerprint acquisitions, is not in the scope of these metrics. This work proposes the use of genuine comparison scores as a measure of variability. It is shown that the publicly available PLUS-MSL-FP dataset exhibits large natural variations which can be used to distinguish between different acquisition sessions. Furthermore, it is showcased that point-cloud (set) based neural networks are promising candidates for processing fingerprint imagery as they provide precise control over the input parameters. Experiments show that point-cloud based neural networks are capable of distinguishing between the different sessions in the PLUS-MSL-FP dataset solely based on FP minutiae locations.
- TextdokumentChallenges of Network Traffic Classification Using Deep Learning in Virtual Networks(INFORMATIK 2022, 2022) Spiekermann,Daniel; Keller,JörgThe increasing number of network-based attacks like denial-of-service and ransomware have become a serious threat in nowadays digital infrastructures. Therefore, the monitoring of network communications and the classification of network packets is a critical process when protecting the environment. Modern techniques like deep learning aim to help the providers when detecting anomalies or attacks by learning details extracted from a network packet or a flow of packets. Most of these models are trained in networks without any kind of virtualisation, especially network virtualisation overlay environments are not investigated in detail. In this paper, we analyse the classification rate of a Convolutional Neural Network (CNN) faced with encapsulated packets. We evaluate this approach with a proof-of-concept based on a binary classification of a self-curated data-set.
- TextdokumentDeep Convolutional Neural Networks for Pose Estimation in Image-Graphics Search(INFORMATIK 2017, 2017) Eberts, Markus; Ulges, AdrianDeep Convolutional Neural Networks (CNNs) have recently been highly successful in various image understanding tasks, ranging from object category recognition over image classification to scene segmentation. We employ CNNs for pose estimation in a cross-modal retrieval system, which -given a photo of an object -allows users to retrieve the best match from a repository of 3D models. As our system is supposed to display retrieved 3D models from the same perspective as the query image (potentially with virtual objects blended over), the pose of the object relative to the camera needs to be estimated. To do so, we study two CNN models. The first is based on end-to-end learning, i.e. a regression neural network directly estimates the pose. The second uses transfer learning with a very deep CNN pre-trained on a large-scale image collection. In quantitative experiments on a set of 3D models and real-world photos of chairs, we compare both models and show that while the end-to-end learning approach performs well on the domain it was trained on (graphics) it suffers from the capability to generalize to a new domain (photos). The transfer learning approach on the other hand handles this domain drift much better, resulting in an average angle deviation from the ground truth angle of about 14 degrees on photos.
- TextdokumentDeep Quality-informed Score Normalization for Privacy-friendly Speaker Recognition in unconstrained Environments(BIOSIG 2017, 2017) Nautsch,Andreas; Steen,Søren Trads; Busch,ChristophIn scenarios that are ambitious to protect sensitive data in compliance with privacy regulations, conventional score normalization utilizing large proportions of speaker cohort data is not feasible for existing technology, since the entire cohort data would need to be stored on each mobile device. Hence, in this work we motivate score normalization utilizing deep neural networks. Considering unconstrained environments, a quality-informed scheme is proposed, normalizing scores depending on sample quality estimates in terms of completeness and signal degradation by noise. Utilizing the conventional PLDA score, comparison i-vectors, and corresponding quality vectors, we aim at mimicking cohort based score normalization optimizing the Cmin llr discrimination criterion. Examining the I4U data sets for the 2012 NIST SRE, an 8.7% relative gain is yielded in a pooled 55-condition scenario with a corresponding condition-averaged relative gain of 6.2% in terms of Cmin llr . Robustness analyses towards sensitivity regarding unseen conditions are conducted, i.e. when conditions comprising lower quality samples are not available during training.
- KonferenzbeitragDeepHyperion: Exploring the Feature Space of Deep Learning-based Systems through Illumination Search(Software Engineering 2023, 2023) Zohdinasab, Tahereh; Riccio, Vincenzo; Gambi, Alessio; Tonella, PaoloWe report about recent research on satisfiability solving for variational domains, originally published in 2022 in the Empirical Software Engineering Journal (EMSE) within the special issue on configurable systems[ Yo22]. Incremental SAT solving is an extension of classic SAT solving that enables solving a set of related SAT problems by identifying and exploiting shared terms. However, using incremental solvers effectively is hard since performance is sensitive to the input order of subterms and results must be tracked manually. This paper translates the ordering problem to an encoding problem and automates the use of incremental solving. We introduce variational SAT solving, which differs from incremental solving by accepting all related problems as a single variational input and returning all results as a single variational output. Variational SAT solving automates the interaction with the incremental solver and enables a method to automatically optimize sharing in the input. We formalize a variational SAT algorithm, construct a prototype variational solver, and perform an empirical analysis on two real-world datasets that applied incremental solvers to software evolution scenarios. We show that the prototype solver scales better for these problems than four off-the-shelf incremental solvers while also automatically tracking individual results.