Deep Convolutional Autoencoder Network Research Articles

Single-Pixel Imaging Based on Deep Learning Enhanced Singular Value Decomposition.

We propose and demonstrate a single-pixel imaging method based on deep learning network enhanced singular value decomposition. The theoretical framework and the experimental implementation are elaborated and compared with the conventional methods based on Hadamard patterns or deep convolutional autoencoder network. Simulation and experimental results show that the proposed approach is capable of reconstructing images with better quality especially under a low sampling ratio down to 3.12%, or with fewer measurements or shorter acquisition time if the image quality is given. We further demonstrate that it has better anti-noise performance by introducing noises in the SPI systems, and we show that it has better generalizability by applying the systems to targets outside the training dataset. We expect that the developed method will find potential applications based on single-pixel imaging beyond the visible regime.

A novel transfer gear life prediction method by the cross-condition health indicator and nested hierarchical binary-valued network

As one of the crucial techniques in prognostics and health management, the long-term remaining useful life (RUL) prediction of gears is essential for maintenance decisions in engineering applications. However, most previous RUL prediction approaches may have unacceptable prediction performance due to insufficient prior life-cycle data and variable operating conditions. In this paper, a novel transfer life prediction method is proposed for gear RUL estimation under different working conditions. First, a health indicator (HI) transfer construction framework, QFMDCAET, is developed for generating cross-condition HIs under various working conditions based on a quadratic function-based multi-scale deep convolutional auto-encoder network and multi-kernel maximum mean discrepancy. With the acquired gear HIs, the nested hierarchical binary-valued network (NHBN) is designed to estimate the gear RUL. In NHBN, a Gumbel function is applied to improve the memory ability of NHBN. Moreover, a nested hierarchical mechanism is also used to fully utilize the sequence information of the cell state in NHBN. The experimental results illustrate the effectiveness of the proposed NHBN in gear RUL prediction, especially for long-term RUL prediction. Lastly, the comparison results indicate that the proposed NHBN-based RUL prediction method is superior to other conventional and advanced RUL prediction methods.

A Depth-Wise Separable U-Net Architecture with Multiscale Filters to Detect Sinkholes

Numerous variants of the basic deep segmentation model—U-Net—have emerged in recent years, achieving reliable performance across different benchmarks. In this paper, we propose an improved version of U-Net with higher performance and reduced complexity. This improvement was achieved by introducing a sparsely connected depth-wise separable block with multiscale filters, enabling the network to capture features of different scales. The use of depth-wise separable convolution significantly reduces the number of trainable parameters, making the training faster, while reducing the risk of overfitting. We used our developed sinkhole dataset and the available benchmark nuclei dataset to assess the proposed model’s performance. Pixel-wise annotation is laborious and requires a great deal of human expertise; therefore, we propose a fully deep convolutional autoencoder network that utilizes the proposed block to automatically annotate the sinkhole dataset. Our segmentation model outperformed the state-of-the-art methods, including U-Net, Attention U-Net, Depth-Separable U-Net, and Inception U-Net, achieving an average improvement of 1.2% and 1.4%, respectively, on the sinkhole and the nuclei datasets, with 94% and 92% accuracy, as well as a reduced training time. It also achieved 83% and 80% intersection-over-union (IoU) on the two datasets, respectively, which is an 11.8% and 9.3% average improvement over the above-mentioned models.

Removing non-resonant background from broadband CARS using a physics-informed neural network.

Broadband coherent anti-Stokes Raman scattering (BCARS) is capable of producing high-quality Raman spectra spanning broad bandwidths, 400-4000 cm-1, with millisecond acquisition times. Raw BCARS spectra, however, are a coherent combination of vibrationally resonant (Raman) and non-resonant (electronic) components that may challenge or degrade chemical analyses. Recently, we demonstrated a deep convolutional autoencoder network, trained on pairs of simulated BCARS-Raman datasets, which could retrieve the Raman signal with high quality under ideal conditions. In this work, we present a new computational system that incorporates experimental measurements of the laser system spectral and temporal properties, combined with simulated susceptibilities. Thus, the neural network learns the mapping between the susceptibility and the measured response for a specific BCARS system. The network is tested on simulated and measured experimental results taken with our BCARS system.

A Deep Convolutional Autoencoder for Automatic Motion Artifact Removal in Electrodermal Activity.

This study aimed to develop a robust and data driven automatic motion artifacts (MA) removal technique from electrodermal activity (EDA) signal. we proposed a deep convolutional autoencoder (DCAE) approach for automatic MA removal in EDA signals. Our model was trained using several publicly available datasets that were collected using a wide variety of stimuli to cause EDA reactions; the sample size was large ([Formula: see text]). We trained and validated our DCAE network using both Gaussian white noise (GWN) and realistic MA data records collected using a novel circuitry in our lab. We further evaluated and compared the performance of our DCAE model with the existing methods on two independent and unseen datasets called Chon lab motion artifact dataset II (CMAD II) and central nervous system oxygen toxicity dataset (CNS-OT). Our DCAE model showed significantly higher signal-to-noise-power-ratio improvement ( SNRimp) and lower mean squared error ( MSE) when compared with that of the three previous methods (averaged [Formula: see text], and MSE = 0.028 on the MA-corrupted data). Moreover, the reconstructed EDAs from the CMAD II dataset had a mean correlation value of 0.78 (statistically significantly higher when compared with other methods) with the reference clean data from the motionless hand, whereas the raw MA-corrupted data had a correlation value of only 0.68. The results presented in the paper indicates that our DCAE can remove MAs with higher intensity where the existing methods fails. Proposed DCAE model can be used to recover a significant amount of otherwise discarded EDA data.

Fast hyperspectral single-pixel imaging via frequency-division multiplexed illumination.

Hyperspectral imaging that detects 3D spectra-spatial information has been used in a wide range of applications. Among reported techniques, multiplexed spectral imaging with a single-pixel detector provides as a photon-efficient and low-cost implementation; however, the previous spectral modulation schemes are mostly complicated and sacrifice the imaging speed. Here, we propose a fast and compact hyperspectral single-pixel imaging technique based on programmable chromatic illumination. A multi-wavelength LED array modulated by independent carriers achieves stable and accurate spectral modulation up to MHz in a frequency-division multiplexed manner, hence allowing the full use of the spatial light modulation speed. Additionally, we propose a multi-channel deep convolutional autoencoder network to reconstruct hyperspectral data from highly-compressed 1D measurement. Experimental reconstructions of 12 spectral channels and 64 × 64 pixels are demonstrated for dynamic imaging at 12 fps image rate. The proposed imaging scheme is highly extensible to a wide spectrum range, and holds potential for portable spectral imagers in low-light or scattering applications.

Spectral-Spatial Deep Support Vector Data Description for Hyperspectral Anomaly Detection

Hyperspectral anomaly detection (HAD) aims to distinguish anomalies from background-by-background modeling. Deep learning has been applied to HAD and achieves promising detection results. However, there exist several issues that need to be addressed: 1) unrealistic Gaussian assumption on the latent representations may limit its application; 2) deep features are not well-suited to anomaly detection due to the separation between feature learning and anomaly detection; 3) lack of adequate exploitation of spectral-spatial features; 4) negative effect caused by spectral band redundancy. In this article, we propose an end-to-end trainable deep one-class classification network for HAD. Specifically, a minimal enclosing hypersphere is trained to involve the deep features of background samples. These background samples are selected by a density clustering-based method. In this way, feature learning and anomaly detection are incorporated into a unified framework. Meanwhile, there is no explicit Gaussian assumption on the background features. Moreover, due to the complementarity of spectral and spatial features, a novel feature fusion strategy is proposed to fuse spectral and spatial features extracted by a two-stream deep convolutional autoencoder network. Finally, a band attention module is used to automatically learn small weights for redundant bands and thus reduce the negative effect caused by redundant bands. Experimental results on five public datasets demonstrate the superiority of the proposed method compared to several state-of-the-art HAD methods in the detection performance.

OutlierNets: Highly Compact Deep Autoencoder Network Architectures for On-Device Acoustic Anomaly Detection.

Human operators often diagnose industrial machinery via anomalous sounds. Given the new advances in the field of machine learning, automated acoustic anomaly detection can lead to reliable maintenance of machinery. However, deep learning-driven anomaly detection methods often require an extensive amount of computational resources prohibiting their deployment in factories. Here we explore a machine-driven design exploration strategy to create OutlierNets, a family of highly compact deep convolutional autoencoder network architectures featuring as few as 686 parameters, model sizes as small as 2.7 KB, and as low as 2.8 million FLOPs, with a detection accuracy matching or exceeding published architectures with as many as 4 million parameters. The architectures are deployed on an Intel Core i5 as well as a ARM Cortex A72 to assess performance on hardware that is likely to be used in industry. Experimental results on the model’s latency show that the OutlierNet architectures can achieve as much as 30× lower latency than published networks.

High imaging quality of Fourier single pixel imaging based on generative adversarial networks at low sampling rate

Single pixel imaging is an innovative imaging scheme using active light to obtain spatial information, which has attracted much attention in the computational imaging field. However, for single pixel imaging, it is a great challenge to find an efficient technique to obtain imaging results with high quality under low sampling conditions. In order to solve this problem, a Fourier single pixel imaging (FSPI) based on a generative adversarial network (GAN) is proposed in this paper. In the proposed GAN model, perceptual loss, pixel and frequency loss are incorporated into the total loss function to better preserve the details of the target. With the help of the GAN model, the FSPI can reconstruct results with high quality at low sampling rate conditions. The numerical simulation and experiment are implemented. Compared with conventional FSPI and FSPI based on a deep convolutional auto-encoder network, the proposed method has a better visual effect and image quality evaluation index. This approach is particularly important to high speed single pixel imaging applications due to its potential for reconstructing the high-quality target image with a low sampling rate.

Image Inpainting Based on Improved Deep Convolutional Auto‐encoder Network

This paper proposes an effective image inpainting method using an improved deep convolutional auto-encoder network. By analogy with exiting methods of image inpainting based on auto-decoders, inpainting methods using the deep convolutional auto-encoder networks are significantly more effective in capturing high-level features than classical methods based on exemplar. However, the inpainted regions would appear blurry and global inconsistency. To alleviate the fuzzy problem, we improved the network model by adding skip connections between mirrored layers in encoder and decoder stacks, so that the generative process of the inpainting area can directly use the low-level features information of the processing image. For making the inpainted result look both more plausible and consistent with its surrounding contexts, the model is trained with a combination of standard pixel-wise reconstruction loss and two adversarial losses which ensures pixel-accurate and local-global contents consistency. With extensive experimental on the ImageNet and Paris Streetview datasets, we demonstrate qualitatively and quantitatively that our approach performs better than state of the art.

Speaker Clustering by Co-Optimizing Deep Representation Learning and Cluster Estimation

Speaker clustering is a task to merge speech segments uttered by the same speaker into a single cluster, which is an effective tool for alleviating the management of massive amount of audio documents. In this paper, we present a work for co-optimizing the two main steps of speaker clustering, namely, feature learning and cluster estimation. In our method, the deep representation feature is learned by a deep convolutional autoencoder network (DCAN), while the cluster estimation is realized by a softmax layer that is combined with the DCAN. We devise an integrated loss function to simultaneously minimize the reconstruction loss (for deep representation learning) and the clustering loss (for cluster estimation). Many state-of-the-art audio features and clustering methods are evaluated on experimental datasets selected from two publicly available speech corpora (the AISHELL-2 and the VoxCeleb1). The results show that the proposed method exceeds other speaker clustering methods in regard to the normalized mutual information (NMI) and the clustering accuracy (CA). Additionally, the proposed deep representation feature outperforms other features that were widely used in previous works, in terms of both NMI and CA.

DeepGhost: real-time computational ghost imaging via deep learning

The potential of random pattern based computational ghost imaging (CGI) for real-time applications has been offset by its long image reconstruction time and inefficient reconstruction of complex diverse scenes. To overcome these problems, we propose a fast image reconstruction framework for CGI, called “DeepGhost”, using deep convolutional autoencoder network to achieve real-time imaging at very low sampling rates (10–20%). By transferring prior-knowledge from STL-10 dataset to physical-data driven network, the proposed framework can reconstruct complex unseen targets with high accuracy. The experimental results show that the proposed method outperforms existing deep learning and state-of-the-art compressed sensing methods used for ghost imaging under similar conditions. The proposed method employs deep architecture with fast computation, and tackles the shortcomings of existing schemes i.e., inappropriate architecture, training on limited data under controlled settings, and employing shallow network for fast computation.

Waveform embedding: Automatic horizon picking with unsupervised deep learning

Picking horizons from seismic images is a fundamental step that could critically impact seismic interpretation quality. We have developed an unsupervised approach, waveform embedding, based on a deep convolutional autoencoder network to learn to transform seismic waveform samples to a latent space in which any waveform can be represented as an embedded vector. The regularizing mechanism of the autoencoder ensures that similar waveform patterns are mapped to embedded vectors with a shorter distance in the latent space. Within a search region, we transform all of the waveform samples to the latent space and compute their corresponding distance to the embedded vector of a control point that is set to the target horizon. We then convert the distance to a horizon probability map that highlights where the horizon is likely to be located. This method can guide horizon picking across lateral discontinuities such as faults, and it is insensitive to noise and lateral distortions. In addition, our unsupervised learning algorithm requires no training labels. We apply our horizon-picking method to multiple 2D/3D examples and obtain results more accurate than the baseline method.

Hyperspectral unmixing using deep convolutional autoencoder

ABSTRACTHyperspectral Unmixing (HU) estimates the combination of endmembers and their corresponding fractional abundances in each of the mixed pixels in the hyperspectral remote sensing image. In this paper, we address the linear unmixing problem with an unsupervised Deep Convolutional Autoencoder network (DCAE). The proposed DCAE is an end-to-end model that consists of two parts: encoder and decoder. First, deep convolutional encoder is employed to extract significant, denoised, non-redundant feature vector. Then one-layer decoder maps the aforementioned feature vector to obtain the endmembers and their abundance percentages. Extensive experiments were carried to evaluate the performance of the proposed method using synthetic and real hyperspectral datasets namely Samson, Cuprite, Urban, and Jasper Ridge. Analyses of the results demonstrate that the proposed DCAE significantly outperforms benchmark unmixing methods even in a noisy environment in terms of both Root Mean Square Error (RMSE), and Mean Square Error (MSE). The achieved results of the proposed DCAE in terms of mean absolute error were 0.0097, 0.001, 0.0141, and 0.0145 for Samson, Cuprite, Urban, and Jasper Ridge datasets respectively.

Deringing and denoising in extremely under-sampled Fourier single pixel imaging.

Undersampling in Fourier single pixel imaging (FSI) is often employed to reduce imaging time for real-time applications. However, the undersampled reconstruction contains ringing artifacts (Gibbs phenomenon) that occur because the high-frequency target information is not recorded. Furthermore, by employing 3-step FSI strategy (reduced measurements with low noise suppression) with a low-grade sensor (i.e., photodiode), this ringing is coupled with noise to produce unwanted artifacts, lowering image quality. To improve the imaging quality of real-time FSI, a fast image reconstruction framework based on deep convolutional autoencoder network (DCAN) is proposed. The network through context learning over FSI artifacts is capable of deringing, denoising, and recovering details in 256 × 256 images. The promising experimental results show that the proposed deep-learning-based FSI outperforms conventional FSI in terms of image quality even at very low sampling rates (1-4%).

Improving Imaging Quality of Real-time Fourier Single-pixel Imaging via Deep Learning.

Fourier single pixel imaging (FSPI) is well known for reconstructing high quality images but only at the cost of long imaging time. For real-time applications, FSPI relies on under-sampled reconstructions, failing to provide high quality images. In order to improve imaging quality of real-time FSPI, a fast image reconstruction framework based on deep learning (DL) is proposed. More specifically, a deep convolutional autoencoder network with symmetric skip connection architecture for real time 96 × 96 imaging at very low sampling rates (5–8%) is employed. The network is trained on a large image set and is able to reconstruct diverse images unseen during training. The promising experimental results show that the proposed FSPI coupled with DL (termed DL-FSPI) outperforms conventional FSPI in terms of image quality at very low sampling rates.

U-DeepHand: FMCW Radar-Based Unsupervised Hand Gesture Feature Learning Using Deep Convolutional Auto-Encoder Network

Recently, although radar sensors have been widely applied for hand gesture recognition (HGR) tasks, conventional radar-based HGR systems still have two major challenges. First, these systems rely on supervised learning approaches to learn gesture features, which normally require a large-scale labeled dataset to address the overfitting problem. However, the acquisition of such dataset is time-consuming. Second, the radar signature of hand movement is often influenced by micromotion caused by other body parts, which leads to distorted motion features, resulting in poor identification accuracy. To overcome these problems, we propose an unsupervised hand gesture feature learning method using the deep convolutional auto-encoder network to analyze hand gesture signal collected by a frequency modulated continuous wave (FMCW) radar sensor. First, via a convolutional encoder sub-network, input radar range profiles are transformed into lower dimensional representations. Then, the representations are expanded to reconstruct the corresponding input profiles by a deconvolutional decoder sub-network. In addition, to investigate the mechanisms of the proposed network and evaluate its performance, we conduct an in-depth study of the feature maps learned from various hand gesture experimental data and evaluate the corresponding classification performance. The results demonstrate that the proposed convolutional auto-encoder network is able to achieve high recognition accuracy with low training sample cost, which outperforms the state-of-the-art hand gesture recognition systems based on transfer learning VGGNet and fully connected-based auto-encoder network.

High-Resolution SAR Image Classification via Deep Convolutional Autoencoders

Synthetic aperture radar (SAR) image classification is a hot topic in the interpretation of SAR images. However, the absence of effective feature representation and the presence of speckle noise in SAR images make classification difficult to handle. In order to overcome these problems, a deep convolutional autoencoder (DCAE) is proposed to extract features and conduct classification automatically. The deep network is composed of eight layers: a convolutional layer to extract texture features, a scale transformation layer to aggregate neighbor information, four layers based on sparse autoencoders to optimize features and classify, and last two layers for postprocessing. Compared with hand-crafted features, the DCAE network provides an automatic method to learn discriminative features from the image. A series of filters is designed as convolutional units to comprise the gray-level cooccurrence matrix and Gabor features together. Scale transformation is conducted to reduce the influence of the noise, which integrates the correlated neighbor pixels. Sparse autoencoders seek better representation of features to match the classifier, since training labels are added to fine-tune the parameters of the networks. Morphological smoothing removes the isolated points of the classification map. The whole network is designed ingeniously, and each part has a contribution to the classification accuracy. The experiments of TerraSAR-X image demonstrate that the DCAE network can extract efficient features and perform better classification result compared with some related algorithms.