Logo des Repositoriums

Auflistung nach Schlagwort "CNN"

1 - 10 von 11
  • Konferenzbeitrag
    Aktuelle Ansätze zum Einsatz von Verfahren der automatisierten Bilderkennung mittels maschinellen Lernens im Bereich des Umweltmonitorings
    (INFORMATIK 2024, 2024) Galle, Christopher
    Die steigende Nachfrage nach präzisen aktuellen Erhebungen des Naturzustands hat die Notwendigkeit neuer Herangehensweisen an die Datenerfassung und -auswertung deutlich gemacht. Die Auswertung von Umweltdaten ist eine zeitaufwändige und ressourcenintensive Aufgabe, die eine erhebliche Beteiligung qualifizierten Personals erfordert. Die Automatisierung dieser oft manuellen Prozesse gestaltete sich über viele Jahre hinweg als herausfordernd. Besonders die Artenbestimmung von Insekten und die Auswertung von Wildkameraaufnahmen im Bereich der Ökologieforschung dienen als Beispiele dafür. In Fangflaschen konservierte Insekten müssen von Fachpersonal identifiziert werden, was aufgrund von Beschädigungen an den Insekten sowie dem Verfall während der Lagerung und Bearbeitung ein zeitaufwendiger und zeitkritischer Prozess ist. Aber nicht nur die Auswertung herkömmlicher Bilder und Proben ist für Anwendungen der automatisierten Bilderkennung interessant, auch nicht-fotografische Bilddaten wie Sonar-, Satelliten- oder spektroskopische Aufnahmen eignen sich dafür. Die Verwendung von Methoden des maschinellen Lernens, insbesondere der Einsatz von Convolutional Neural Networks, hat sich hier in vielen Bereichen als äußerst hilfreich erwiesen. Die Verfügbarkeit geeigneter Trainingsdaten stellt jedoch weiterhin ein großes Problem dar, für das häufig individuelle Lösungsansätze gefunden werden müssen
  • Konferenzbeitrag
    An anomaly detection approach for backdoored neural networks: face recognition as a case study
    (BIOSIG 2022, 2022) Alexander Unnervik and Sébastien Marcel
    Backdoor attacks allow an attacker to embed functionality jeopardizing proper behavior of any algorithm, machine learning or not. This hidden functionality can remain inactive for normal use of the algorithm until activated by the attacker. Given how stealthy backdoor attacks are, consequences of these backdoors could be disastrous if such networks were to be deployed for applications as critical as border or access control. In this paper, we propose a novel backdoored network detection method based on the principle of anomaly detection, involving access to the clean part of the training data and the trained network.We highlight its promising potential when considering various triggers, locations and identity pairs, without the need to make any assumptions on the nature of the backdoor and its setup. We test our method on a novel dataset of backdoored networks and report detectability results with perfect scores.
  • Textdokument
    Automated Architecture-Modeling for Convolutional Neural Networks
    (BTW 2019 – Workshopband, 2019) Duong, Manh Khoi
    Tuning hyperparameters can be very counterintuitive and misleading, yet it plays a big (or even the biggest) part in many machine learning algorithms. For instance, finding the right architecture for an artificial neural network (ANN) can also be seen as a hyperparameter e.g. number of convolutional layers, number of fully connected layers etc. Tuning these can be done manually or by techniques such as grid search or random search. Even then finding optimal hyperparameters seems to be impossible. This paper tries to counter this problem by using bayesian optimization, which finds optimal parameters, including the right architecture for ANNs. In our case, a histological image dataset was used to classify breast cancer into stages.
  • Zeitschriftenartikel
    Automatic Classification of Bloodstains with Deep Learning Methods
    (KI - Künstliche Intelligenz: Vol. 36, No. 2, 2022) Bergman, Tommy; Klöden, Martin; Dreßler, Jan; Labudde, Dirk
    The classification of detected bloodstains into predetermined categories is a crucial component of the so-called bloodstain pattern analysis. As in other forensic disciplines, deep learning methods may help to reduce human subjectivity within this process, may increase the classification accuracy, shorten the calculation time and thus, enable high-throughput analysis. In this work, an approach is presented in which a convolutional neural network (Inception v3) was trained from 965 drip stains (passive origin) and 1595 blood spatters (active origin). The trained CNN was evaluated with a test data set consisting of 366 images of drip stains and blood spatters. The success rate was 99.73% which suggests that neural networks could also be used to automatically classify other classes of bloodstain patterns to speed up the investigation process in the future.
  • Konferenzbeitrag
    End-to-end Off-angle Iris Recognition Using CNN Based Iris Segmentation
    (BIOSIG 2020 - Proceedings of the 19th International Conference of the Biometrics Special Interest Group, 2020) Jalilian, Ehsaneddin; Karakaya, Mahmut; Uhl, Andreas
    While deep learning techniques are increasingly becoming a tool of choice for iris segmentation, yet there is no comprehensive recognition framework dedicated for off-angle iris recognition using such modules. In this work, we investigate the effect of different gaze-angles on the CNN based off-angle iris segmentations, and their recognition performance, introducing an improvement scheme to compensate for some segmentation degradations caused by the off-angle distortions. Also, we propose an off-angle parameterization algorithm to re-project the off-angle images back to frontal view. Taking benefit of these, we further investigate if: (i) improving the segmentation outputs and/or correcting the iris images before or after the segmentation, can compensate for off-angle distortions, or (ii) the generalization capability of the network can be improved, by training it on iris images of different gaze-angles. In each experimental step, segmentation accuracy and the recognition performance are evaluated, and the results are analyzed and compared.
  • Textdokument
    Explainable Diagnosis of COVID-19 from Chest X-ray Images via CNNs
    (SKILL 2021, 2021) Arkan, Emre; Beckert, Jan Malte
    This work demonstrates how Convolutional Neural Networks ( CNN s) can be used to identify signs of COVID-19 from Chest X-rays (CXR s) and discusses the challenges of deep learning with small datasets. In order to validate the model’s performance, two novel explanation methods LIME and Grad-CAM are explored. Additionally, they serve to further increase users’ confidence in specific classifications. Since the explanation results revealed model biases, additional preprocessing mechanisms were explored: A U-Net-based lung segmenter is introduced to the preprocessing pipeline, which masks all non-lung parts of the CXRs images. Subsequently, the segmentation and non-segmentation results were evaluated with regard to both their performance metrics and interpreted explanation results.
  • Konferenzbeitrag
    Hyper-Parameter Search for Convolutional Neural Networks - An Evolutionary Approach
    (SKILL 2018 - Studierendenkonferenz Informatik, 2018) Bibaeva, Victoria
    Convolutional neural networks is one of the most popular neural network classes within the deep learning research area. Due to their specific architecture they are widely used to solve such challenging tasks as image and speech recognition, video analysis etc. The architecture itself is defined by a number of (hyper-)parameters that have major impact on the recognition rate. Although much significant progress has been made to improve the performance of convolutional networks, the typical hyper-parameter search is done manually, taking therefore a long time and likely to disregard some very good values. This paper solves the problem by proposing two different evolutionary algorithms for automated hyper-parameter search in convolutional architectures. It will be shown that in case of image recognition these algorithms are capable of finding architectures with nearly state of the art performance automatically, sparing the scientists from much tedious effort.
  • Konferenzbeitrag
    Incorporation of Extra Pseudo Labels for CNN-based Gait Recognition
    (BIOSIG 2022, 2022) Daigo Muramatsu, Kousuke Moriwaki
    CNN is a major model used for image-based recognition tasks, including gait recognition, and many CNN-based network structures and/or learning frameworks have been proposed. Among them, we focus on approaches that use multiple labels for learning, typified by multi-task learning. These approaches are sometimes used to improve the accuracy of the main task by incorporating extra labels associated with sub-tasks. The incorporated labels for learning are usually selected from real tasks heuristically; for example, gender and/or age labels are incorporated together with subject identity labels.We take a different approach and consider a virtual task as a sub-task, and incorporate pseudo output labels together with labels associated with the main task and/or real task. In this paper, we focus on a gait-based person recognition task as the main task, and we discuss the effectiveness of virtual tasks with different pseudo labels for construction of a CNN-based gait feature extractor.
  • Konferenzbeitrag
    On the detection of morphing attacks generated by GANs
    (BIOSIG 2022, 2022) Laurent Colbois and Sébastien Marcel
    Recent works have demonstrated the feasibility of GAN-based morphing attacks that reach similar success rates as more traditional landmark-based methods. This new type of “deep” morphs might require the development of new adequate detectors to protect face recognition systems. We explore simple deep morph detection baselines based on spectral features and LBP histograms features, as well as on CNN models, both in the intra-dataset and cross-dataset case. We observe that simple LBP-based systems are already quite accurate in the intra-dataset setting, but struggle with generalization, a phenomenon that is partially mitigated by fusing together several of those systems at score-level.We conclude that a pretrained ResNet effective for GAN image detection is the most effective overall, reaching close to perfect accuracy. We note however that LBP-based systems maintain a level of interest : additionally to their lower computational requirements and increased interpretability with respect to CNNs, LBP+ResNet fusions sometimes also showcase increased performance versus ResNet-only, hinting that LBP-based systems can focus on meaningful signal that is not necessarily picked up by the CNN detector.
  • Konferenzbeitrag
    Rotation Tolerant Finger Vein Recognition using CNNs
    (BIOSIG 2021 - Proceedings of the 20th International Conference of the Biometrics Special Interest Group, 2021) Promegger, Bernhard; Wimmer, Georg; Uhl, Andreas
    Finger vein recognition deals with the recognition of subjects based on their venous pattern within the fingers. The majority of the available systems acquire the vein pattern using only a single camera. Such systems are susceptible to misplacements of the finger during acquisition, in particular longitudinal finger rotation poses a severe problem. Besides some hardware based approaches that try to avoid the misplacement in the first place, there are several software based solutions to counter fight longitudinal finger rotation. All of them use classical hand-crafted features. This work presents a novel approach to make CNNs robust to longitudinal finger rotation by training CNNs using finger vein images from varying perspectives.