Rozdziały

Stały URI dla kolekcjihttp://hdl.handle.net/11652/4775

Przeglądaj

Filtry

20237

Ustawienia

Ma przypisane plikisearch.filters.applied.operator.equals: TakTematsearch.filters.applied.operator.equals: search.filters.subject.deep learning

collection.search.results.head

Teraz wyświetlane 1 - 7 z 7
  • Pozycja
    Mixing Synthetic and Real-world Datasets Strategy for Improved Generalization of the CNN
    (Wydawnictwo Politechniki Łódzkiej, 2023) Młodzikowski, Kamil; Belter, Dominik
    In this paper, we deal with the problem of supervised training neural networks with an insufficient number of real-world training examples. We propose a method that at the beginning trains the neural network using a relatively simple synthetic dataset. In the following epochs, we add more challenging and real-life images to the training dataset. We compare the proposed strategy with other methods of using artificial and real-world datasets for training the neural network. The obtained results show that the proposed strategy allows for obtaining the neural network with higher generalization capabilities than competitive methods.
  • Pozycja
    Improving RGB-D Visual Odometry with Depth Learned from a Better Sensor’s Output
    (Wydawnictwo Politechniki Łódzkiej, 2023) Kostusiak, Aleksander
    This paper compares the results obtained from an indoor Visual Odometry (VO) system with RGB-D images provided by a Kinect v1 camera against those achieved by a VO with enhanced depth channel. For this purpose, we have used two classic image inpainting methods and a deeplearning approach for scene depth estimation employing Kinect v2 depth maps as reference data. The ability to enhance lower-quality data is crucial to reduce the cost of VO applications because higher-quality information can be infused through deep learning in systems using budget sensors.
  • Pozycja
    A New Approach to Learning of 3D Characteristic Points for Vehicle Pose Estimation
    (Wydawnictwo Politechniki Łódzkiej, 2023) Nowak, Tomasz; Skrzypczyński, Piotr
    This article discusses the challenges of estimating the pose of a vehicle from monocular images in an uncontrolled environment. We propose a new neural network architecture that learns 3D characteristic points of vehicles from image crops and coordinates of 2D keypoints on images. To facilitate supervised training of this network, we pre-process the ApolloCar3D dataset to obtain labelled 3D characteristic points of different car models. We evaluate our approach on the ApolloCar3D benchmark and demonstrate results competitive to state-of-the-art methods.
  • Pozycja
    Towards Ontology-Driven Verification of Car Claims Settlement
    (Wydawnictwo Politechniki Łódzkiej, 2023) Pancerz, Krzysztof; Wolski, Jacek
    In the paper, we outline an intelligent tool enabling the users to automatize the process of verification of the car claims settlement. Two data sources power the tool. The first one is the source of car images in which damaged elements are recognized. The second one is the source of PDF files in which cost estimates are extracted. The designed ontology of car repair, described in the paper, is used both in the pre-processing step and in recognition of a typical situations.
  • Pozycja
    Text-to-music Models and Their Evaluation Methods
    (Wydawnictwo Politechniki Łódzkiej, 2023) Modrzejewski, Mateusz; Rokita, Przemysław
    Text-to-music models are a very recent approach to generative music, allowing to generate music based on an abstract, rich description input in natural language. In this paper, we propose guidelines for evaluation in text-to-music models, highlighting the need for musical insight and clear descriptions of perceptual quality upon investigating the metrics of currently developed approaches. We also present a critical analysis of the capabilities and evaluation methods of the pioneering text-to-music models.
  • Pozycja
    Spotting Advertisements from Above: Billboard Detection and Segmentation in UAV Imagery
    (Wydawnictwo Politechniki Łódzkiej, 2023) Ptak, Bartosz; Dominiak, Jan; Kraft, Marek
    In this work, deep-learning methods were researched for billboard detection in urban environments. Billboards are one of the adversarial visual pollutants occurring in cities, causing over-saturation of visual stimulation. Due to this, we develop an algorithm that helps in the analysis and management of urban space. We utilise near real-time object detection methods to detect and segment them on images registered by unmanned aerial vehicles (UAVs). Research is based on recent algorithms from the YOLO family with modified heads for the instance segmentation task. We gathered images and prepared hand-annotated labels for training and evaluation purposes of deep learning approaches. We reached the mAP@0.5 metric of 0.61 for detection and 0.60 for segmentation, enabling us to develop smart city applications.
  • Pozycja
    Fault Diagnosis in a Squirrel-Cage Induction Motor Using Thermal Imaging
    (Wydawnictwo Politechniki Łódzkiej, 2023) Piechocki, Mateusz; Kraft, Marek; Pajchrowski, Tomasz
    Fault diagnosis is a vivid topic in industrial applications or intelligent building solutions. One of the well-established techniques involves the measurement and analysis of current signals. However, this method has several significant drawbacks, such as the inability to inspect during machinery operation or the lack of precise information on the malfunction location. This article proposes a non-invasive method for squirrel-cage induction motor’s state classification and fault diagnosis. The approach is based on thermal image analysis that utilizes a compact convolution neural network. In addition, the gathered and annotated image set, which consists of thermal images with 640 x 512 pixels resolution, is presented.