Convolutional Autoencoder Architecture Research Articles

A Comparative Study of Cross-Device Finger Vein Recognition Using Classical and Deep Learning Approaches

Finger vein recognition is gaining popularity in the field of biometrics, yet the inter-operability of finger vein patterns has received limited attention. This study aims to fill this gap by introducing a cross-device finger vein dataset and evaluating the performance of finger vein recognition across devices using a classical method, a convolutional neural network, and our proposed patch-based convolutional auto-encoder (CAE). The findings emphasise the importance of standardisation of finger vein recognition, similar to that of fingerprints or irises, crucial for achieving inter-operability. Despite the inherent challenges of cross-device recognition, the proposed CAE architecture in this study demonstrates promising results in finger vein recognition, particularly in the context of cross-device comparisons.

Pattern recognition and signal parameters extraction using machine learning methods

Machine learning methods can be used for signal processing in different cases of physics research. A convolutional neural network was developed for the task of pulse counting in particle detectors for high energy physics. For the extraction of signal parameters was used a network with convolutional autoencoder architecture and a subsequent result reconstruction algorithm was developed and applied. A convolutional neural network was also developed for seismic studies with the task of identifying different events in seismograms. All of the algorithms with their architecture, input and output are presented and discussed.

Finite element model-informed deep learning for equivalent force estimation and full-field response calculation

The measured data in structural health monitoring are discrete in space, which means that the full-field load distribution or structural responses cannot be captured. Conventional deep-learning-based data expansion methods require a huge volume of training datasets, and they tend to be overfitted because constraints are insufficient. To address these issues, a physics-informed neural network (PINN) is developed in this study by estimating an arbitrarily distributed load to an equivalent load. Subsequently, a finite element model is used to reconstruct structural responses. The proposed PINN is divided into connected trainable data model and untrainable physics-driven model. The data-driven model in the proposed PINN is a convolutional autoencoder architecture, while the physics-informed model incorporates the impulse response matrix derived from the time-invariant system. The autoencoder is fed with the monitored data, and an equivalent load is outputted. Consecutively, the impulse response function in the physics-informed model is used to calculate the responses at the sensor positions to the equivalent load. The difference between the measured and predicted responses is formulated as the loss function to train the PINN. In this way, the physics-driven model is integrated into the data model to guide its training direction in a novel self-supervised manner, thereby avoiding the overfitting effects in network training. In addition, the unmeasured and randomly-distributed load can be estimated automatically. The proposed PINN is applied to a numerical continuous beam and a laboratory-tested three-storey frame, where only a few sensors are required for response reconstruction. Results suggest that the PINN can accurately reconstruct the full-field structural responses via the estimated equivalent load and applies to complex conditions with random and unknown load distributions.

Fully Automatic Approach for Smoke Tracking Based on Deep Image Quality Enhancement and Adaptive Level Set Model

In recent decades, the need for advanced systems with good precision, low cost, and high-time response for wildfires and smoke detection and monitoring has become an absolute necessity. In this paper, we propose a novel, fast, and autonomous approach for denoising and tracking smoke in video sequences captured from a camera in motion. The proposed method is based mainly on two stages: the first one is a reconstruction and denoising path with a novel lightweight convolutional autoencoder architecture. The second stage is a specific scheme designated for smoke tracking, and it consists of the following: first, the foreground frames are extracted with the HSV color model and textural features of smoke; second, possible false detections of smoke regions are eliminated with image processing technique and last smoke contours detection is performed with an adaptive nonlinear level set. The obtained experimental results exposed in this paper show the potential of the proposed approach and prove its efficiency in smoke video denoising and tracking with a minimized number of false negative regions and good detection rates.

Unsteady-state turbulent flow field predictions with a convolutional autoencoder architecture

<abstract> <p>Traditional numerical methods, such as computational fluid dynamics (CFD), demand large computational resources and memory for modeling fluid dynamic systems. Hence, deep learning (DL) and, specifically Convolutional Neural Networks (CNN) autoencoders have resulted in accurate tools to obtain approximations of the streamwise and vertical velocities and pressure fields, when stationary flows are considered. The novelty of this paper consists of predicting the future instants from an initial one with a CNN autoencoder architecture when an unsteady flow is considered. Two neural models are proposed: The former predicts the future instants on the basis of an initial sample and the latter approximates the initial sample. The inputs of the CNNs take the signed distance function (SDF) and the flow region channel (FRC), and, for the representation of the temporal evolution, the previous CFD sample is added. To increment the amount of training data of the second neural model, a data augmentation technique based on the similarity principle for fluid dynamics is implemented. As a result, low absolute error rates are obtained in the prediction of the first samples near the shapes surfaces. Even in the most advanced time instants, the prediction of the vortices zone is quite reliable. 62.12 and 9000 speed-up ratios are achieved by the predictions of the first and second neural models, respectively, compared to the computational cost regarded by the CFD simulations.</p> </abstract>

A Machine Learning Approach for Atherosclerosis Detection in Murine Models

The early detection of atherosclerosis has been the interest of researchers in order to prevent diseases progression. And, most of cases diagnosed with atherosclerosis in late stages resulted of subjective ways of diagnosis in the healthcare sector. Consequently, in order to help cardiologists diagnose atherosclerosis in early stages with an objective, fast and accurate way, we are proposing a machine learning model based on autoencoders that enables atherosclerosis detection from MRI images of murine subjects. This novel way of automated system consists of applying various image processing techniques on the input MRI images. Then, after training, testing and validation it will be capable of classifying the images as atherosclerotic or not based on a specific threshold for the reconstruction error calculated from the autoencoders' output. An autoencoder is a feed-forward neural network that has its input neurons equal to the output neuron. The classification works by comparing the reconstructed images to the original input images and evaluating loss between them. Since the autoencoder is trained on healthy images, reconstruction error of the healthy images would be low while that of atherosclerotic subjects would be higher. By setting a threshold for the loss, we can classify the images as healthy or atherosclerotic. The pre-processing of these images was made using a Block Matching 3D (BM3D) filter to remove the noise in the images prior to application of a Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance the contrast. The next step included introducing the dataset into the autoencoder to start training on the healthy images, after increasing the images number with augmentation. The results showed a reconstruction loss of 0.018 while using the stacked architecture of the autoencoder and 0.0366 when using the convolutional autoencoder architecture.

Learning two-phase microstructure evolution using neural operators and autoencoder architectures

Phase-field modeling is an effective but computationally expensive method for capturing the mesoscale morphological and microstructure evolution in materials. Hence, fast and generalizable surrogate models are needed to alleviate the cost of computationally taxing processes such as in optimization and design of materials. The intrinsic discontinuous nature of the physical phenomena incurred by the presence of sharp phase boundaries makes the training of the surrogate model cumbersome. We develop a framework that integrates a convolutional autoencoder architecture with a deep neural operator (DeepONet) to learn the dynamic evolution of a two-phase mixture and accelerate time-to-solution in predicting the microstructure evolution. We utilize the convolutional autoencoder to provide a compact representation of the microstructure data in a low-dimensional latent space. After DeepONet is trained in the latent space, it can be used to replace the high-fidelity phase-field numerical solver in interpolation tasks or to accelerate the numerical solver in extrapolation tasks.

Deep-Learning-Based Signal Enhancement of Low-Resolution Accelerometer for Fall Detection Systems

In the last two decades, fall detection (FD) systems have been developed as a popular assistive technology. To support long-term FD services, various power-saving strategies have been implemented. Among them, a reduced sampling rate is a common approach for an energy-efficient system in the real world. However, the performance of FD systems is diminished owing to low-resolution (LR) accelerometer signals. To improve the detection accuracy with LR accelerometer signals, several technical challenges must be considered, including mismatch of effective features and the degradation effects. In this work, a deep-learning-based accelerometer signal enhancement (ASE) model is proposed as a front-end processor to help typical LR-FD systems achieve better detection performance. The proposed ASE model based on a deep denoising convolutional autoencoder architecture reconstructs high-resolution (HR) signals from the LR signals by learning the relationship between the LR and HR signals. The results show that the FD system using support vector machine (SVM) and the proposed ASE model at an extremely low sampling rate (sampling rate < 2 Hz) achieved 97.34% and 90.52% accuracies in the SisFall and FallAllD data sets, respectively, while those without ASE models only achieved 95.92% and 87.47% accuracies in the SisFall and FallAllD data sets, respectively. The results also demonstrate that the proposed ASE mode can be suitably combined with deep-learning-based FD systems.

Residual Learning to Integrate Neural Network and Physics-Based Models for Improved Production Prediction in Unconventional Reservoirs

Summary The flow and transport processes that take place during hydrocarbon production from hydraulically fractured unconventional reservoirs are not well understood. As a result, current simulators cannot provide reliable predictions of the production behavior in the field. In addition to imperfect physics, the prediction errors can be caused by the inability to conveniently integrate important field data, such as well logs, drilling, and completion parameters, into existing physical models. A neural network (NN) model is developed to learn the (residual) errors in simulation-based production prediction as a funcation of input parameters of an unconventional well. Once trained, the NN model augments the physics-based predictions by adding the learned reiodual to predict the production response of a new well. To learn the discrepancy between the simulated and observed production data, the NN model is trained using a labeled dataset consisting of the prediction errors (as labels) and the corresponding input parameters (features), such as formation, completion, and fluid properties. During training, a mapping is identified from the input parameters to their respective prediction errors. To facilitate the residual learning, first a convolutional autoencoder architecture is used to map the simulated and observed production responses to a low-dimensional latent space. This step is followed by a regression model that learns the mapping between the collected field parameters and the corresponding latent space representation of the prediction errors. The two steps are included in a single NN architecture and trained simultaneously. The proposed residual learning method is designed to compensate for prediction errors originating from a combination of imperfect representation of the physics and inaccurate simulation inputs, including uncertain descriptions of the reservoir and fracture properties. The performance of the proposed residual learning approach is evaluated using synthetic data as well as a field case study from the Bakken play in North Dakota.

A deep learning framework for characterization of genotype data.

Dimensionality reduction is a data transformation technique widely used in various fields of genomics research. The application of dimensionality reduction to genotype data is known to capture genetic similarity between individuals, and is used for visualization of genetic variation, identification of population structure as well as ancestry mapping. Among frequently used methods are principal component analysis, which is a linear transform that often misses more fine-scale structures, and neighbor-graph based methods which focus on local relationships rather than large-scale patterns. Deep learning models are a type of nonlinear machine learning method in which the features used in data transformation are decided by the model in a data-driven manner, rather than by the researcher, and have been shown to present a promising alternative to traditional statistical methods for various applications in omics research. In this study, we propose a deep learning model based on a convolutional autoencoder architecture for dimensionality reduction of genotype data. Using a highly diverse cohort of human samples, we demonstrate that the model can identify population clusters and provide richer visual information in comparison to principal component analysis, while preserving global geometry to a higher extent than t-SNE and UMAP, yielding results that are comparable to an alternative deep learning approach based on variational autoencoders. We also discuss the use of the methodology for more general characterization of genotype data, showing that it preserves spatial properties in the form of decay of linkage disequilibrium with distance along the genome and demonstrating its use as a genetic clustering method, comparing results to the ADMIXTURE software frequently used in population genetic studies.

Deep Compression for Dense Point Cloud Maps

Many modern robotics applications rely on 3D maps of the environment. Due to the large memory requirements of dense 3D maps, compression techniques are often necessary to store or transmit 3D maps efficiently. In this work, we investigate the problem of compressing dense 3D point cloud maps such as those obtained from an autonomous vehicle in large outdoor environments. We tackle the problem by learning a set of local feature descriptors from which the point cloud can be reconstructed efficiently and effectively. We propose a novel deep convolutional autoencoder architecture that directly operates on the points themselves so that we avoid voxelization. Additionally, we propose a deconvolution operator to upsample point clouds, which allows us to decompress to an arbitrary density. Our experiments show that our learned compression achieves better reconstructions at the same bit rate compared to other state-of-the-art compression algorithms. We furthermore demonstrate that our approach generalizes well to different LiDAR sensors. For example, networks learned on maps generated from KITTI point clouds still achieve state-of-the-art compression results for maps generated from nuScences point clouds.

End-to-End Image Steganography Using Deep Convolutional Autoencoders

Image steganography is used to hide a secret image inside a cover image in plain sight. Traditionally, the secret data is converted into binary bits and the cover image is manipulated statistically to embed the secret binary bits. Overloading the cover image may lead to distortions and the secret information may become visible. Hence the hiding capacity of the traditional methods are limited. In this paper, a light-weight yet simple deep convolutional autoencoder architecture is proposed to embed a secret image inside a cover image as well as to extract the embedded secret image from the stego image. The proposed method is evaluated using three datasets - COCO, CelebA and ImageNet. Peak Signal-to-Noise Ratio, hiding capacity and imperceptibility results on the test set are used to measure the performance. The proposed method has been evaluated using various images including Lena, airplane, baboon and peppers and compared against other traditional image steganography methods. The experimental results have demonstrated that the proposed method has higher hiding capacity, security and robustness, and imperceptibility performances than other deep learning image steganography methods.

Deep convolutional neural network-based anomaly detection for organ classification in gastric X-ray examination

Aim:The aim of this study was to determine whether our deep convolutional neural network-based anomaly detection model can distinguish differences in esophagus images and stomach images obtained from gastric X-ray examinations. Methods:A total of 6012 subjects were analyzed as our study subjects. Since the number of esophagus X-ray images is much smaller than the number of gastric X-ray images taken in X-ray examinations, we took an anomaly detection approach to realize the task of organ classification. We constructed a deep autoencoding gaussian mixture model (DAGMM) with a convolutional autoencoder architecture. The trained model can produce an anomaly score for a given test X-ray image. For comparison, the original DAGMM, AnoGAN, and a One-Class Support Vector Machine (OCSVM) that were trained with features obtained by a pre-trained Inception-v3 network were used. Results:Sensitivity, specificity, and the calculated harmonic mean of the proposed method were 0.956, 0.980, and 0.968, respectively. Those of the original DAGMM were 0.932, 0.883, and 0.907, respectively. Those of AnoGAN were 0.835, 0.833, and 0.834, respectively, and those of OCSVM were 0.932, 0.935, and 0.934, respectively. Experimental results showed the effectiveness of the proposed method for an organ classification task. Conclusion:Our deep convolutional neural network-based anomaly detection model has shown the potential for clinical use in organ classification.

A Similarity Searching System for Biological Phenotype Images Using Deep Convolutional Encoder-decoder Architecture

Background: The BLAST (Basic Local Alignment Search Tool) algorithm has been widely used for sequence similarity searching. Analogously, the public phenotype images must be efficiently retrieved using biological images as queries and identify the phenotype with high similarity. Due to the accumulation of genotype-phenotype-mapping data, a system of searching for similar phenotypes is not available due to the bottleneck of image processing. Objective: In this study, we focus on the identification of similar query phenotypic images by searching the biological phenotype database, including information about loss-of-function and gain-of-function. Methods: We propose a deep convolutional autoencoder architecture to segment the biological phenotypic images and develop a phenotype retrieval system to enable a better understanding of genotype–phenotype correlation. Results: This study shows how deep convolutional autoencoder architecture can be trained on images from biological phenotypes to achieve state-of-the-art performance in a phenotypic images retrieval system. Conclusion: Taken together, the phenotype analysis system can provide further information on the correlation between genotype and phenotype. Additionally, it is obvious that the neural network model of image segmentation and the phenotype retrieval system is equally suitable for any species, which has enough phenotype images to train the neural network.

Energy Compaction-Based Image Compression Using Convolutional AutoEncoder

Image compression has been an important research topic for many decades. Recently, deep learning has achieved great success in many computer vision tasks, and its use in image compression has gradually been increasing. In this paper, we present an energy compaction-based image compression architecture using a convolutional autoencoder (CAE) to achieve high coding efficiency. Our main contributions include three aspects: 1) we propose a CAE architecture for image compression by decomposing it into several down(up)sampling operations; 2) for our CAE architecture, we offer a mathematical analysis on the energy compaction property and we are the first work to propose a normalized coding gain metric in neural networks, which can act as a measurement of compression capability; 3) based on the coding gain metric, we propose an energy compaction-based bit allocation method, which adds a regularizer to the loss function during the training stage to help the CAE maximize the coding gain and achieve high compression efficiency. The experimental results demonstrate our proposed method outperforms BPG (HEVC-intra), in terms of the MS-SSIM quality metric. Additionally, we achieve better performance in comparison with existing bit allocation methods, and provide higher coding efficiency compared with state-of-the-art learning compression methods at high bit rates.

Adversarial Training for Speech Super-Resolution

Speech super-resolution or speech bandwidth expansion aims to upsample a given speech signal by generating the missing high-frequency content. In this paper, we propose a deep neural network approach exploiting the adversarial training ideas that have been shown effective in image super-resolution. Specifically, our proposed network follows the generative adversarial networks setup, where the generator network uses a convolutional autoencoder architecture with one-dimensional convolution kernels to generate high-frequency log-power spectra from the low-frequency log-power spectra of the input speech. We propose to use both the reconstruction loss and the adversarial loss for training, and we employ a recent regularization method that penalizes the gradient norms of the discriminator to stabilize the training. We compare our proposed approach with two state-of-the-art neural network baselines and evaluate these methods with both objective speech quality measures and subjective perceptual and intelligibility tests. Results show that our proposed method outperforms both baselines in terms of both objective and subjective evaluations. To gain insights of the network architecture, we analyze key parameters of the proposed network including the number of layers, the number of convolution kernels, and the relative weight of the reconstruction and adversarial losses. Besides, we analyze the computational complexity of our method and the baselines and discuss ways for phase estimation. We further develop a noise-resilient version of the proposed approach by training the network with noisy speech inputs. Objective evaluation validates the noise-resilient property on unseen noise types.

Depth-based subgraph convolutional auto-encoder for network representation learning

Network representation learning (NRL) aims to map vertices of a network into a low-dimensional space which preserves the network structure and its inherent properties. Most existing methods for network representation adopt shallow models which have relatively limited capacity to capture highly non-linear network structures, resulting in sub-optimal network representations. Therefore, it is nontrivial to explore how to effectively capture highly non-linear network structure and preserve the global and local structure in NRL. To solve this problem, in this paper we propose a new graph convolutional autoencoder architecture based on a depth-based representation of graph structure, referred to as the depth-based subgraph convolutional autoencoder (DS-CAE), which integrates both the global topological and local connectivity structures within a graph. Our idea is to first decompose a graph into a family of K-layer expansion subgraphs rooted at each vertex aimed at better capturing long-range vertex inter-dependencies. Then a set of convolution filters slide over the entire sets of subgraphs of a vertex to extract the local structural connectivity information. This is analogous to the standard convolution operation on grid data. In contrast to most existing models for unsupervised learning on graph-structured data, our model can capture highly non-linear structure by simultaneously integrating node features and network structure into network representation learning. This significantly improves the predictive performance on a number of benchmark datasets.

A study of deep convolutional auto-encoders for anomaly detection in videos

The detection of anomalous behaviors in automated video surveillance is a recurrent topic in recent computer vision research. Depending on the application field, anomalies can present different characteristics and challenges. Convolutional Neural Networks have achieved the state-of-the-art performance for object recognition in recent years, since they learn features automatically during the training process. From the anomaly detection perspective, the Convolutional Autoencoder (CAE) is an interesting choice, since it captures the 2D structure in image sequences during the learning process. This work uses a CAE in the anomaly detection context, by applying the reconstruction error of each frame as an anomaly score. By exploring the CAE architecture, we also propose a method for aggregating high-level spatial and temporal features with the input frames and investigate how they affect the CAE performance. An easy-to-use measure of video spatial complexity was devised and correlated with the classification performance of the CAE. The proposed methods were evaluated by means of several experiments with public-domain datasets. The promising results support further research in this area.