Examinando por Materia "Convolutional neural network"

Mostrando 1 - 2 de 2
  • No hay miniatura disponible
    Ítem
    Aplicación de técnicas de clusterización para la clasificación de música dance electrónica
    (Universidad EAFIT, 2023) Murillo Martínez, Carlos Alberto; Alunno, Marco; Martínez Vargas, Juan David
    Audio processing is one of the essential tasks for a data scientist, and audio analysis has applications in a diverse range of fields, such as medicine, telecommunications, improving sound quality in music production, and even military applications (filtering suspicious or terrorist audio). This project aims to use hard clustering techniques (such as k-means or k-nearest neighbor) and soft clustering techniques (such as fuzzy clustering) to classify input songs using different metrics. The classification methods will be used to segment previously processed input audios and obtain a sample of representative segments of the songs, determining their similarity with other songs of the same genre. Another technique that has proven effective for audio classification is convolutional neural networks (CNNs), which have been used in a wide range of fields. In the music field, they have been used to classify violin bowing techniques [1] and even detect potential heart problems using heartbeat sounds [2]. In this project, we will use this technique up to the point of feature extraction, and then use classical classification techniques to determine which group a section of a song belongs to.
  • No hay miniatura disponible
    Ítem
    FocusNET : an autofocusing learning‐based model for digital lensless holographic microscopy
    (Universidad EAFIT, 2023) Montoya Zuluaga, Manuel; Trujillo Anaya, Carlos Alejandro; Lopera Acosta, María Josef
    This paper reports on a convolutional neural network (CNN) – based regression model, called FocusNET, to predict the accurate reconstruction distance of raw holograms in Digital Lensless Holographic Microscopy (DLHM). This proposal provides a physical-mathematical formulation to extend its use to different DLHM setups than the optical and geometrical conditions utilized for recording the training dataset; this unique feature is tested by applying the proposal to holograms of diverse samples recorded with different DLHM setups. Additionally, a comparison between FocusNET and conventional autofocusing methods in terms of processing times and accuracy is provided. Although the proposed method predicts reconstruction distances with approximately 54 µm standard deviation, accurate information about the samples in the validation dataset is still retrieved. When compared to a method that utilizes a stack of reconstructions to find the best focal plane, FocusNET performs 600 times faster, as no hologram reconstruction is needed. When implemented in batches, the network can achieve up to a 1200-fold reduction in processing time, depending on the number of holograms to be processed. The training and validation datasets, and the code implementations, are hosted on a public GitHub repository that can be freely accessed.