Figure 1 - uploaded by Herminarto Nugroho
Content may be subject to copyright.
Source publication
![Figure 1. Comparison between PCA and Autoencoder [15]](profile/Herminarto-Nugroho/publication/340049776/figure/fig1/AS:871003374747649@1584674795907/Comparison-between-PCA-and-Autoencoder-15_Q320.jpg)
This paper proposes a fully convolutional variational autoencoder (VAE) for features extraction from a large-scale dataset of fire images. The dataset will be used to train the deep learning algorithm to detect fire and smoke. The features extraction is used to tackle the curse of dimensionality, which is the common issue in training deep learning...
Citations
... [23]. The autoencoder is used to find new input representations without losing too much information and the input can be reconstructed [24]. Inputs in autoencoder can be reconstructed effectively with minimum reconstruction error [25]. ...
... The Model Classifier can be trained and tested using the output. An autoencoder was used to extract the dimensions, after which they were converted into features for 27,26,25,24,23,22,21,20,19,18,14,10, and 6 neurons. The Autoencoder (AE) model is configured with the activation function at the bottleneck, encoder 1, encoder 2, decoder 1, and encoder 2 using Relu, the kernel initializer used is Random Normal with a standard deviation of 0.01, the initializer bias used is Zeros, and linear activation is used on the output layers. ...
... The information contained in these features can be used as input for secondary processes such as classification, segmentation, and object detection. Nugroho et al. (2020) discussed the important information that is retained by the VAE model when it compresses the input. They also suggested that the features from the latent vector can provide suitable information and can substitute the original images. ...
Tidal flats are among the ecologically richest areas of the world where sediment composition (e.g. median grain size and silt content) and the macrozoobenthic presence play an important role in the health of the ecosystem. Regular monitoring of environmental and ecological variables is essential for sustainable management of the area. While monitoring based on field sampling is very time-consuming, the predictive performance of these variables using satellite images is low due to the spectral homogeneity over these regions. We tested a novel approach that uses features from a variational autoencoder (VAE) model to enhance the predictive performance of remote sensing images for environmental and ecological variables of tidal flats. The model was trained using the Sentinel-2 spectral bands to reproduce the input images, and during this process, the VAE model represents important information on the tidal flats within its layer structure. The information in the layers of the trained model was extracted to form features with identical spatial coverage to the spectral bands. The features and the spectral bands together form the input to random forest models to predict field observations of the sediment characteristics such as median grain size and silt content, as well as the macrozoobenthic biomass and species richness. The maximum prediction accuracy of feature-based maps was close to 62% for the sediment characteristics and 37% for benthic fauna indices. The encoded features improved the prediction accuracy of the random forest regressor model by 15% points on average in comparison to using just the spectral bands. Our method enhances the predictive performance of remote sensing, in particular the spatiotemporal dynamics in median grain size and silt content of the sediment thereby contributing to better-informed management of coastal ecosystems.
... Thus, autoencoders can learn complex data representation, which can be employed to describe the original data in a lower dimensionality [34]. Fig. 1, obtained from Ref. [122], shows the difference between linear and nonlinear dimensionality reduction methods. ...
... At this point, the loss Fig. 1. Difference between linear and nonlinear dimensionality reduction methods (obtained from Ref. [122]). function should be minimized for training the model. ...
Computational fluid dynamics (CFD) models have been used for the simulation of urban airflow and pollutant dispersion, due to their capability to capture different length scales and turbulence nature of the flow field. However, their high computational costs prevent them from being used for (near) real-time simulations, long-term predictions, and simulations with dynamic boundary conditions. Reduced-order models (ROMs) are proposed as reliable alternatives to CFD approaches to solve the mentioned issues. This article aims to comprehensively review the state-of-the-art application of different methodologies to develop a non-intrusive ROM (NIROM) for predicting urban airflow and pollutant dispersion. Developing such models comprises two steps: dimensionality reduction and computing the feature dynamics of the reduced space. Various methodologies, with the focus on machine learning algorithms, are proposed for the mentioned stages, while their capabilities and limitations are discussed. Furthermore, different approaches are introduced to overcome the issue of the physical interpretation of these models. Also, several methods are proposed to make the models suitable for being used in long-term predictions and multi-query problems (i.e., considering changes in boundary conditions).
... Both PCA and the β-VAE can be used for dimensionality reduction and feature extraction [8,9]. The choice of suitable features, however, requires assessment metrics. ...
The usability of machine learning approaches for the development of in-situ process monitoring, automated anomaly detection and quality assurance for the selective laser melting (SLM) process receives currently increasing attention. For a given set of real machine data we compare two established methods, principal component analysis (PCA) and ß-variational autoencoder (ß-VAE), for their applicability in exploratory data analysis and anomaly detection. We introduce a PCA-based unsupervised feature extraction algorithm, which allows for root cause analysis of process anomalies. The ß-VAE enables a slightly more compact dimensionality reduction; we consider it an option for automated process monitoring systems.
... Autoencoder vs. PCA dimension reduction task[8] ...
There may be parts in the images with different features that are not desired. These
undesired parts might be caused by aging of a photograph or these parts could include
an object which is wanted to be removed from the image. Removing these parts and
inpainting them with visually realistic features is both a research area and a commercial
problem. There are many specialized tools for this purpose and use of these tools may
also require professional knowledge.
The aim of the Image Inpainting with Deep Learning project is to examine the different
structures in the pictures that fill these regions that are extracted from the pictures.
For this purpose, many different work types were divided into groups and sample
studies from these groups were compared. Inpainting methods for review purposes,
are divided into two as traditional methods and deep methods. Although many
methods were mentioned during our study only a few are chosen. Navier-strokes and
fast-marching methods were chosen for comparison as example of traditional methods.
EdgeConnect, Generative Contextual, Generative Multi-column Convolutional Neural
Network and Deep Image Prior methods, which were found to give successful results
among deep methods, were selected to be examined in detaial and implemented. In
addition to all these, two novel methods are implemented which are inspired from all
methods we examined throughout this project. Results of these two methods are also
explained and examined in this project.
As examples of deep methods, our two methods, EdgeConnect and Generative
Contextual are two GAN based studies which use deep artificial neural networks that
work adversarial to each other. There are two stages in the EdgeConnect method and
these two stages use the mentioned GAN structure. In the first stage, corrupted image
is taken as input to the model and edge information is predicted, and in the second
stage, it is desired to produce an image whose missing region is filled by using the
both edge information and masked image. Also, in Generative Contextual study there
are two stages that use GAN aswell. In the first stage, maske area is inpainted roughlt,
and in the second stage, improvement operations are carried out on the created image
by using a layer that aims to obtain contextual information.
Another deep image inpainting study chosen is the Generative Multi-column CNN
method, which uses convolutional layers in parallel. This method passes the input
image through parallel CNN structures using filters of different kernel sizes. Then,
these structures, which provide different levels of information on the image, are
brought together. And as a result, the inpainting process is completed.
Finally, another method that we have examined in detail is the Deep Image Prior
method. This is a method that performs the inpainting process without the need to
train a model using any dataset. This method, which can also be used in applications such as super-resolution and image denoising, obtains iterative results by using various
CNN structures.
Pytorch and Tensorflow deep learning libraries were used during our study. Our studies
were run on the graphics card, especially since serious processing power was required
for training and testing of deep learning methods. As a result of our project, different
studies that we tested were compared and presented with PSNR and SSIM values and
example images.
Aquaculture plays a significant role in both economic development and food production. Maintaining an ecological environment with good water quality is essential to ensure the production efficiency and quality of aquaculture. Effective management of water quality can prevent abnormal conditions and contribute significantly to food security. Detecting anomalies in the aquaculture environment is crucial to ensure that the environment is maintained correctly to meet healthy and proper requirements for fish farming. This article focuses on the use of deep learning techniques to detect anomalies in water quality data in the aquaculture environment. Four deep learning anomaly detection techniques, including Autoencoder, Variational Autoencoder, Long-Short Term Memory Autoencoder, and Spectral-Residual Convolutional Neural Network, were analysed using multiple real-world sensor datasets collected from IoT aquaculture systems. Extensive experiments were conducted for temperature, dissolved oxygen, and pH parameters, and the evaluation analysis revealed that the Long-Short Term Memory Autoencoder anomaly detection method showed promising results in detecting anomalies for the temperature and oxygen datasets, while the Spectral-Residual Convolutional Neural Network demonstrated the best performance on the pH datasets.
The Internet of Things (IoT) refers to a system of interconnected, internet-connected devices and sensors that allows the collection and dissemination of data. The data provided by these sensors may include outliers or exhibit anomalous behavior as a result of attack activities or device failure, for example. However, the majority of existing outlier detection algorithms rely on labeled data, which is frequently hard to obtain in the IoT domain. More crucially, the IoT’s data volume is continually increasing, necessitating the requirement for predicting and identifying the classes of future data. In this study, we propose an unsupervised technique based on a deep Variational Auto-Encoder (VAE) to detect outliers in IoT data by leveraging the characteristic of the reconstruction ability and the low-dimensional representation of the input data’s latent variables of the VAE. First, the input data are standardized. Then, we employ the VAE to find a reconstructed output representation from the low-dimensional representation of the latent variables of the input data. Finally, the reconstruction error between the original observation and the reconstructed one is used as an outlier score. Our model was trained only using normal data with no labels in an unsupervised manner and evaluated using Statlog (Landsat Satellite) dataset. The unsupervised model achieved promising and comparable results with the state-of-the-art outlier detection schemes with a precision of ≈90% and an F1 score of 79%.
