Fakulta elektrotechniky a informatiky / Faculty of Electrical Engineering and Informatics
Stálý URI pro tuto komunituhttps://hdl.handle.net/10195/3847
Práce obhájené před rokem 2008 jsou uloženy pouze v kolekci Vysokoškolské kvalifikační práce
Procházet
8 výsledky
Search Results
Konferenční objektpeer-reviewedpostprint Omezený přístup ECG Hearbeat Classification Based on Multi-scale Convolutional Neural Networks(Springer Nature Switzerland AG, 2023) Rozinek, Ondřej; Doležel, PetrClinical applications require automating ECG signal processing and classification. This paper investigates the impact of multiscale input filtering techniques and feature map blocks on the performance of CNN models for ECG classification. We conducted an ablation study using the AbnormalHeartbeat dataset, with 606 instances of ECG time series divided into five classes. We compared five multiscale input filtering techniques and four multiscale feature map blocks against a base model and non-multiscale input. Results showed that the combination of mean filter for multiscale input and residual connections for multiscale block achieved the highest accuracy of 64.47%. Residual connections were consistently effective across different filtering techniques, highlighting their potential to enhance CNN model performance for ECG classification. These findings can guide the design of future CNN models for ECG classification tasks, with further experimentation needed for optimal combinations in specific applications.Konferenční objektpeer-reviewedpostprint Omezený přístup Automated Dataset Enhancement Using GAN for Assessment of Degree of Degradation around Scribe(IEEE (Institute of Electrical and Electronics Engineers), 2023) Doležel, Petr; Pakosta, Marek; Rozsívalová, Veronika; Štursa, DominikCoil coating is a method of applying an organic coating material to a rolled metal strip substrate in a continuous automated process. It is used to provide a high quality, durable finish to a variety of surfaces. The degradation resistance of coil-coated materials is assessed according to European Standard EN 13523-8 by exposing a coil-coated test specimen to a salt fog at a defined temperature for a defined period of time. After this process, a sample is tested according to the International Organisation for Standardisation ISO 4628 standard to determine the degree of degradation. In this study, a GAN-based technique for automated training set enhancement is proposed to assess the degree of degradation around a scribe. The presented technique is capable of enhancing a manually generated dataset of images with synthetic samples to help refine the performance of the area degradation detector.Konferenční objektpeer-reviewedpostprint Omezený přístup Multi-Scale Neural Model for Tool-Narayanaswamy-Moynihan Model Parameter Extraction(Springer Nature Switzerland AG, 2023) Pakosta, Marek; Doležel, Petr; Svoboda, Roman; Baruque Zanon, BrunoGlass transitions are an important phenomenon in amor phous materials with potential for various applications. The Tool-Narayanaswamy-Moynihan (TNM) model is a widely used empirical model that describes the enthalpy relaxation behavior of these materials. However, determining the appropriate values for its parameters can be challeng ing. To address this issue, a multi-scale convolutional neural model is pro posed that can accurately predict the TNM parameters directly from the set of differential scanning calorimetry curves, experimentally measured using the sample of the considered amorphous material. The resulting Mean Absolute Error of the model over the test set is found to be 0.0252, indicating a high level of accuracy. Overall, the proposed neural model has the potential to become a valuable tool for practical application of the TNM model in the glass industry and related fields.Konferenční objektpeer-reviewedpostprint Omezený přístup Triple Parallel LSTM Networks for Classifying the Gait Disorders Using Kinect Camera and Robot Platform During the Clinical Examination(IEEE (Institute of Electrical and Electronics Engineers), 2023) Shayestegan, Mohsen; Zálabský, Tomáš; Mareš, JanThis paper presents a new methodology for the data processing and classification method for gait disorders, which is observed with a Kinect camera. The study of gait and motion stability in gait disorders is one of the most interesting research areas in the field. The patient and the physician must monitor the progress of the rehabilitation process before and after surgery to obtain an objective view of the rehabilitation process. In this study, the patient is scanned with the Kinect camera placed on a mobile robotic platform. For feature extraction and feature analysis, the exercises (three walking exercises) frames are collected and saved in data folders. This study uses 84 measurements of 37 patients with complex observations based on the physician's opinion in a clinical setting to address classification problems. In the analysis of gait disorders, motion data play an essential role. Furthermore, it reduces the selection of helpful body features for assessing gait disorders. The proposed system uses a key-point detector that computes body landmarks and classifies gait disorders using triple-parallel long short-term memory (LSTM) networks. The present study demonstrates the success of the method in classification evaluation when combined with the state-of-the-art pose estimation method. Around 81 percent accuracy was achieved for given sets of individuals using velocity-based, angle-based, and position-based features.Konferenční objektpeer-reviewedpostprint (accepted version) Otevřený přístup Comparison of Floating-point Representations for the Efficient Implementation of Machine Learning Algorithms(IEEE, 2022) Mishra, Saras Mani; Tiwari, Ankita; Shekhawat, Hanumant Singh; Guha, Prithwijit; Trivedi, Gaurav; Pidanič, Jan; Němec, ZdeněkSmart systems are enabled by artificial intelligence (AI), which is realized using machine learning (ML) techniques. ML algorithms are implemented in the hardware using fixedpoint, integer, and floating-point representations. The performance of hardware implementation gets impacted due to very small or large values because of their limited word size. To overcome this limitation, various floating-point representations are employed, such as IEEE754, posit, bfloat16 etc. Moreover, for the efficient implementation of ML algorithms, one of the most intuitive solutions is to use a suitable number system. As we know, multiply and add (MAC), divider and square root units are the most common building blocks of various ML algorithms. Therefore, in this paper, we present a comparative study of hardware implementations of these units based on bfloat16 and posit number representations. It is observed that posit based implementations perform 1.50x better in terms of accuracy, but consume 1.51x more hardware resources as compared to bfloat16 based realizations. Thus, as per the trade-off between accuracy and resource utilization, it can be stated that the bfloat16 number representation may be preferred over other existing number representations in the hardware implementations of ML algorithms.Konferenční objektpeer-reviewedpostprint (accepted version) Omezený přístup Classification of Polymers Based on the Degree of Their Transparency in SWIR Spectrum(Springer Nature Switzerland AG, 2022) Štursa, Dominik; Kopecký, Dušan; Roleček, Jiří; Doležel, Petr; Baruque Zanon, BrunoDetection, classification and sorting of polymeric particles is a common task required in recycling industry. In the proposed work, an innovative method for detection of polymeric particles and their classification is introduced. The method is based on evaluation of images of polymeric particles, obtained from short-wavelength infrared (SWIR) camera, by convolutional neural network (CNN). Compared to conventionally used spectroscopes or hyper-spectral imaging, this method utilizes single wavelength (1 050 nm) and a degree of polymer transparency serves as the main descriptor. Five different polymers (ABS, ABS-T, Nylon, PETG, PLA) in form of regular blocks (size 15 × 15 × 0.3 mm) were used in the experiment. In total 203 images (size 288 × 288 px) were prepared for CNN training and 67 for testing. Scalable ASP U-Net was tested in 6 combinations and their outputs were compared. According to used intersection over union metrics over all outputs, the topology with 64 filters and depth of 3 exhibited the best results.Konferenční objektpeer-reviewedpostprint (accepted version) Otevřený přístup Spectral Classification of Microplastics using Neural Networks: Pilot Feasibility Study(SciTePress - Science and Technology Publications, 2022) Doležel, Petr; Roleček, Jiří; Honc, Daniel; Štursa, Dominik; Baruque Zanon, BrunoMicroplastics, i.e. synthetic polymers that have particle size smaller than 5 mm, are emerging pollutants that are widespread in the environment. In order to monitor environmental pollution by microplastics, it is necessary to have available rapid screening techniques, which provide the accurate information about the quality (type of polymer) and quantity (amount). Spectroscopy is an indispensable method, if precise classification of individual polymers in microplastics is required. In order to contribute to the topic of autonomous spectra matching when using spectroscopy, we decided to demonstrate the quality and efficiency of neural networks. We adopted three neural network architectures, and we tested them for application to spectra matching. In order to keep our study transparent, we use publicly available dataset of FTIR spectra. Furthermore, we performed a deep statistical analysis of all the architectures performance and efficiency to show the suitability of neural networks for spectra matching. The results presented at the end of this article indicated the overall suitability of the selected neural network architectures for spectra matching in microplastics classification.Konferenční objektpeer-reviewedpostprint Omezený přístup One Step Deep Learning Approach to Grasp Detection in Robotics(Springer Science and Business Media, 2021) Doležel, Petr; Štursa, Dominik; Honc, DanielGrasp point detection is a necessary ability to handle for industrial robots. In recent years, various deep learning-based techniques for robotic grasping have been introduced. To follow this trend, we introduce a convolutional neural network-based approach for model-free one step method for grasp point detection. This method provides all feasible grasp points suitable for parallel grippers, based on a single RGB image of the scene. A case study, which shows the outstanding accuracy of the presented approach as well as its acceptable response time, is presented at the end of this contribution.