93 research outputs found
Διαφάνεια και προσβασιμότητα στην επεξεργασία ήχου: υβριδικές τεχνικές βαθιάς μάθησης CNN-LST για τον διαχωρισμό φωνής στη μουσική
No full textΠτυχιακή εργασία--Πανεπιστήμιο Μακεδονίας, Θεσσαλονίκη, 2024.Η βιβλιοθήκη διαθέτει αντίτυπο της πτυχιακής μόνο σε ηλεκτρονική μορφή.Vocal separation from audio mixtures is a complex and critical task within
audio signal processing, with significant applications in music remixing, track
creation, and music information retrieval. This study presents a hybrid deep
learning model combining Convolutional Neural Networks (CNNs) and Long
Short-Term Memory (LSTM) networks, designed to effectively isolate vocals
from intricate audio mixtures.
The research leverages the MUSDB18 dataset, a recognized benchmark
for music source separation. Extensive data preprocessing was employed, including the conversion of audio files into spectrograms via Short-Time Fourier
Transform (STFT) and data augmentation techniques such as pitch shifting
and time stretching to enhance the model’s generalization.
The CNN-LSTM model was trained using the Adam optimizer with Mean
Squared Error (MSE) as the loss function. Hyperparameter optimization
was conducted using Optuna, systematically tuning critical parameters to
maximize model performance. Post-processing techniques, including mask
refinement and dynamic range compression, were applied to ensure quality
vocal separations.
To facilitate deployment in resource-constrained environments, dynamic
quantization was applied using the Open Neural Network Exchange (ONNX)
format. This approach reduced the model size and improved inference speed
while maintaining accuracy.
This research advances music source separation by providing an easy-tofollow pipeline for vocal isolation. The methodologies and results presented
offer valuable insights for real-time audio processing applications, paving the
way for broader adoption in various industries
Enhancing Deep Learning Model Explainability in Brain Tumor Datasets Using Post-Heuristic Approaches
No full textThe application of deep learning models in medical diagnosis has showcased considerable efficacy in recent years. Nevertheless, a notable limitation involves the inherent lack of explainability during decision-making processes. This study addresses such a constraint by enhancing the interpretability robustness. The primary focus is directed towards refining the explanations generated by the LIME Library and LIME image explainer. This is achieved through post-processing mechanisms based on scenario-specific rules. Multiple experiments have been conducted using publicly accessible datasets related to brain tumor detection. Our proposed post-heuristic approach demonstrates significant advancements, yielding more robust and concrete results in the context of medical diagnosis
Post-Heuristic Cancer Segmentation Refinement over MRI Images and Deep Learning Models
No full textLately, deep learning methods have greatly improved the accuracy of brain-tumor segmentation, yet slice-wise inconsistencies still limit reliable use in clinical practice. While volume-aware 3D convolutional networks achieve high accuracy, their memory footprint and inference time may limit clinical adoption. This study proposes a resource-conscious pipeline for lower-grade-glioma delineation in axial FLAIR MRI that combines a 2D Attention U-Net with a guided post-processing refinement step. Two segmentation backbones, a vanilla U-Net and an Attention U-Net, are trained on 110 TCGA-LGG axial FLAIR patient volumes under various loss functions and activation functions. The Attention U-Net, optimized with Dice loss, delivers the strongest baseline, achieving a mean Intersection-over-Union (mIoU) of 0.857. To mitigate slice-wise inconsistencies inherent to 2D models, a White-Area Overlap (WAO) voting mechanism quantifies the tumor footprint shared by neighboring slices. The WAO curve is smoothed with a Gaussian filter to locate its peak, after which a percentile-based heuristic selectively relabels the most ambiguous softmax pixels. Cohort-level analysis shows that removing merely 0.1–0.3% of ambiguous low-confidence pixels lifts the post-processing mIoU above the baseline while improving segmentation for two-thirds of patients. The proposed refinement strategy holds great potential for further improvement, offering a practical route for integrating deep learning segmentation into routine clinical workflows with minimal computational overhead
Enhancing Deep Learning Model Explainability in Brain Tumor Datasets using Post-Heuristic Approaches
No full textThe application of deep learning models in medical diagnosis has showcased
considerable efficacy in recent years. Nevertheless, a notable limitation
involves the inherent lack of explainability during decision-making processes.
This study addresses such a constraint, by enhancing the interpretability
robustness. The primary focus is directed towards refining the explanations
generated by the LIME Library and LIME image explainer. This is achieved
throuhg post-processing mechanisms, based on scenario-specific rules. Multiple
experiments have been conducted using publicly accessible datasets related to
brain tumor detection. Our proposed post-heuristic approach demonstrates
significant advancements, yielding more robust and concrete results, in the
context of medical diagnosis
Τεχνικές μάθησης μερικής επίβλεψης για λήψη αποφάσεων
No full textSemi-supervised learning (SSL) is a class of supervised learning tasks and techniques that also exploits the unlabeled data for training. SSL significantly reduces labeling related costs and is able to handle large data sets. The primary objective is the extraction of robust inference rules.Decision support systems (DSSs) who utilize SSL have significant advantages. Only a small amount of labelled data is required for the initialization. Then, new (unlabeled) data can be utilized and improve system’s performance. Thus, the DSS is continuously adopted to new conditions, with minimum effort. Techniques which are cost effective and easily adopted to dynamic systems, can be beneficial for many practical applications. Such applications fields are: (a) industrial assembly lines monitoring, (b) sea border surveillance, (c) elders’ falls detection, (d) transportation tunnels inspection, (e) concrete foundation piles defect recognition, (f) commercial sector companies financial assessment and (g) image advanced filtering for cultural heritage applications.Ο όρος μάθηση με μερική επίβλεψη αναφέρεται σε ένα ευρύ πεδίο τεχνικών μηχανικής μάθησης, οι οποίες χρησιμοποιούν τα μη τιτλοφορημένα δεδομένα για να εξάγουν επιπλέον ωφέλιμη πληροφορία. Η μερική επίβλεψη αντιμετωπίζει προβλήματα που σχετίζονται με την επεξεργασία και την αξιοποίηση μεγάλου όγκου δεδομένων και τα όποια κόστη σχετίζονται με αυτά (π.χ. χρόνος επεξεργασίας, ανθρώπινα λάθη). Απώτερος σκοπός είναι η ασφαλή εξαγωγή συμπερασμάτων, κανόνων ή προτάσεων. Τα μοντέλα λήψης απόφασης που χρησιμοποιούν τεχνικές μερικής μάθησης έχουν ποικίλα πλεονεκτήματα. Σε πρώτη φάση, χρειάζονται μικρό πλήθος τιτλοφορημένων δεδομένων για την αρχικοποίηση τους. Στη συνέχεια, τα νέα δεδομένα που θα εμφανιστούν αξιοποιούνται και τροποποιούν κατάλληλα το μοντέλο. Ως εκ τούτου, έχουμε ένα συνεχώς εξελισσόμενο μοντέλο λήψης αποφάσεων, με την ελάχιστη δυνατή προσπάθεια.Τεχνικές που προσαρμόζονται εύκολα και οικονομικά είναι οι κατεξοχήν κατάλληλες για τον έλεγχο συστημάτων, στα οποία παρατηρούνται συχνές αλλαγές στον τρόπο λειτουργίας. Ενδεικτικά πεδία εφαρμογής εφαρμογής ευέλικτων συστημάτων υποστήριξης λήψης αποφάσεων με μερική μάθηση είναι: η επίβλεψη γραμμών παραγωγής, η επιτήρηση θαλάσσιων συνόρων, η φροντίδα ηλικιωμένων, η εκτίμηση χρηματοπιστωτικού κινδύνου, ο έλεγχος για δομικές ατέλειες και η διαφύλαξη της πολιτιστικής κληρονομιάς
Smart parking solutions for urban areas
No full textFinding a parking place in a busy city centre is often a frustrating task for many drivers; time and fuel are wasted in the quest for a vacant spot and traffic in the area increases due to the slow moving vehicles circling around. In this paper, we present the results of a survey on the needs of drivers from parking infrastructures from a smart services perspective. As smart parking systems are becoming a necessity in today's urban areas, we discuss the latest trends in parking availability monitoring, parking reservation and dynamic pricing schemes. We also examine how these schemes can be integrated forming technologically advanced parking infrastructures whose aim is to benefit both the drivers and the parking operators alike
Outlier detection in maritime environments using AIS data and deep recurrent architectures
No full textA methodology based on deep recurrent models for maritime surveillance, over
publicly available Automatic Identification System (AIS) data, is presented in
this paper. The setup employs a deep Recurrent Neural Network (RNN)-based
model, for encoding and reconstructing the observed ships' motion patterns. Our
approach is based on a thresholding mechanism, over the calculated errors
between observed and reconstructed motion patterns of maritime vessels.
Specifically, a deep-learning framework, i.e. an encoder-decoder architecture,
is trained using the observed motion patterns, enabling the models to learn and
predict the expected trajectory, which will be compared to the effective ones.
Our models, particularly the bidirectional GRU with recurrent dropouts,
showcased superior performance in capturing the temporal dynamics of maritime
data, illustrating the potential of deep learning to enhance maritime
surveillance capabilities. Our work lays a solid foundation for future research
in this domain, highlighting a path toward improved maritime safety through the
innovative application of technology.Comment: Presented in PETRA '24 The PErvasive Technologies Related to
Assistive Environments Conference June 26--28, 2024 Crete, Greec
- …
