Extract Subnetworks Research Articles

Multiframe blind restoration with image quality prior

In dynamic environments, such as ground-based optic observations for space targets, anisotropic turbulence and random noise tend to produce different discriminate image degradations at every moment. These degradations severely reduce the quality of the target images and make restoration of these images very difficult. However, an inconsistent degradation implies that there is complementary information in such images. In this paper, a ranking network (Ranknet) is first proposed to ensure that input sequences have a consistent distribution upon degradation and squeeze the spatial distribution of the sample set. Then, an extraction-refinement neural network (ERnet) is proposed to extract complementary features and blindly reconstruct a clean image of the observation target. In ERnet, an extraction subnetwork (EN) uses 3D convolutions to extract discriminate features from multiframe input sequences, and a refinement subnetwork (RN) based on 2D convolution restores clean images by refining the effective features. In addition, a spatial–temporal attention module (STAM) is devoted to enhancing features through utilization of the high-quality features. Experimental results on the restorations of space target images and motion blur images confirm the superior performance of ERnet as compared with other state-of-the-art methods.

Asymmetric two-stream network for screen content image quality assessment based on region features

With the latest popularization of remote desktop software and online work, screen content images (SCIs) are increasingly being contacted by people. Many image quality assessment (IQA) methods based on deep learning have been proposed for SCIs. However, most IQA methods for SCIs pay less attention to regional characteristics. An asymmetric two-stream network based on region features is proposed for SCI objective quality assessment, in which features are derived from a channel separation feature extraction (CSFE) sub-network and a spatial analysis feature extraction (SAFE) sub-network. In the CSFE sub-network, three channel separation (CS) modules are utilized to strengthen the ability for capturing deepthwise features in the pictorial regions. In the SAFE sub-network, two spatial analysis (SA) module are applied to focus on the structural information in the spatial domain for the textual regions. Two asymmetric sub-networks can efficiently extract the corresponding region features by separately considering the pictorial and textual regions of SCIs. In addition, an adaptive global weighting strategy is introduced to estimate the quality for SCIs. The experimental results on the screen image quality assessment database show that the proposed model is superior to the other state-of-the-art SCI quality assessment methods.

Two-Branch Attention Adversarial Domain Adaptation Network for Hyperspectral Image Classification

Recent studies have shown that deep domain adaptation (DA) techniques have good performance on cross-domain hyperspectral image (HSI) classification problems. However, most existing deep HSI DA approaches directly use deep networks to extract features from the data, which ignores the detailed information of HSI in spectral and spatial dimensions. To effectively exploit the spectral–spatial joint information for DA of HSIs, we propose a two-branch attention adversarial DA (TAADA) network in this article. In the TAADA network, a two-branch feature extraction (TBFE) subnetwork is first designed as a generator to extract the attention-based spectral–spatial features. Then, a discriminator based on two classifiers with the multilayer FC-BN-ReLU-Dropout structure is constructed. Based on adversarial learning between the generator and the discriminator, the ability of discriminative feature extraction and cross-domain classification is improved simultaneously. Finally, the TAADA network can adjust the distribution between the source and target domains and extract domain-invariant features. Experimental results on three cross-scene HSI classification tasks show that our proposed TAADA outperforms some existing DA methods.

Multispectral Pansharpening Based on Multisequence Convolutional Recurrent Neural Network

Multispectral (MS) pansharpening is defined as the fusion of spatial information in panchromatic (PAN) image and spectral information in MS image. In this work, we propose a MS pansharpening based on multi-sequence convolutional recurrent neural network (MCRNN). The proposed MCRNN contains two sub-networks (shallow feature extraction sub-network and deep feature fusion sub-network). In the shallow feature extraction sub-network, PAN and MS images are superimposed in the spectral dimension as multi-sequence data. A convolutional neural network (CNN) based on residual learning is then used to obtain the feature maps from multi-sequence data. In the deep feature fusion sub-network, since MS and PAN images are highly correlated, a convolutional recurrent neural network (ConvGRU) belonging to RNN is used to model adjacent and across-band relationships between these feature maps to capture the local and global correlations of the features in different bands. The global average pooling is then performed on the output results to yield the pansharpening result. Several datasets are tested at reduced and full resolution experiments, the experimental results show that the performance of the proposed MCRNN is superior to the traditional pansharpening methods. The source code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/HZC-1998/Multi-Sequence-Convolutional-Recurrent-Network-for-Pansharpening</uri> .

SCANet: A Spatial and Channel Attention based Network for Partial-to-Partial Point Cloud Registration

Point cloud registration plays an essential role in many areas, such as computer vision and robotics. However, traditional feature-based registration requires handcrafted descriptors for various scenarios, which is of low efficiency and flexibility; ICP and its locally optimal variants are sensitive to initialization, while globally optimal methods are of high computational time to overcome noise, outliers, and partial overlap. Learning-based registration can automatically and flexibly learn shape representation for different objects, but existing methods are of either low efficiency or low precision, and poorly perform in partial-to-partial point cloud registration. Thus, we present a simple spatial and channel attention based network, named SCANet, for partial-to-partial point cloud registration. A spatial self-attention aggregation (SSA) module is applied in a feature extraction sub-network to efficiently make use of the inter and global information of each point cloud in different levels, while a channel cross-attention regression (CCR) module is adopted in a pose estimation sub-network for information interaction between two input global feature vectors, enhancing relevant information and suppressing redundant information. Experimental results show that our SCANet achieves state-of-the-art performances in both accuracy and efficiency compared to existing non-deep learning and learning-based methods on partial visibility with Gaussian noise. Our source code is available at the project website https://github.com/zhouruqin/SCANet.

Transcriptional landscape of cellular networks reveal interactions driving the dormancy mechanisms in cancer

Primary cancer cells exert unique capacity to disseminate and nestle in distant organs. Once seeded in secondary sites, cancer cells may enter a dormant state, becoming resistant to current treatment approaches, and they remain silent until they reactivate and cause overt metastases. To illuminate the complex mechanisms of cancer dormancy, 10 transcriptomic datasets from the literature enabling 21 dormancy–cancer comparisons were mapped on protein–protein interaction networks and gene-regulatory networks to extract subnetworks that are enriched in significantly deregulated genes. The genes appearing in the subnetworks and significantly upregulated in dormancy with respect to proliferative state were scored and filtered across all comparisons, leading to a dormancy–interaction network for the first time in the literature, which includes 139 genes and 1974 interactions. The dormancy interaction network will contribute to the elucidation of cellular mechanisms orchestrating cancer dormancy, paving the way for improvements in the diagnosis and treatment of metastatic cancer.

Multilabel CNN-Based Hybrid Learning Metric for Pedestrian Reidentification

Pedestrian reidentification has recently emerged as a hot topic that attains considerable attention since it can be applied to many potential applications in the surveillance system. However, high-accuracy pedestrian reidentification is a stimulating research problem because of variations in viewpoints, color, light, and other reasons. This work addresses the interferences and improves pedestrian reidentification accuracy by proposing two novel algorithms, pedestrian multilabel learning, and investigating hybrid learning metrics. First, unlike the existing models, we construct the identification framework using two subnetworks, namely, part detection subnetwork and feature extraction subnetwork, to obtain pedestrian attributes and low-level feature scores, respectively. Then, a hybrid learning metric that combines pedestrian attributes and low-level feature scores is proposed. Both low-level features and pedestrian attributes are utilized, thus enhancing the identification rate. Our simulation results on both datasets, i.e., CUHK03 and VIPeR, reveal that the identification rate is improved compared to the existing pedestrian reidentification methods.

DeepHBSP: A Deep Learning Framework for Predicting Human Blood-Secretory Proteins Using Transfer Learning

The identification of blood-secretory proteins and the detection of protein biomarkers in the blood have an important clinical application value. Existing methods for predicting blood-secretory proteins are mainly based on traditional machine learning algorithms, and heavily rely on annotated protein features. Unlike traditional machine learning algorithms, deep learning algorithms can automatically learn better feature representations from raw data, and are expected to be more promising to predict blood-secretory proteins. We present a novel deep learning model (DeepHBSP) combined with transfer learning by integrating a binary classification network and a ranking network to identify blood-secretory proteins from the amino acid sequence information alone. The loss function of DeepHBSP in the training step is designed to apply descriptive loss and compactness loss to the binary classification network and the ranking network, respectively. The feature extraction subnetwork of DeepHBSP is composed of a multi-lane capsule network. Additionally, transfer learning is used to train a highly accurate generalized model with small samples of blood-secretory proteins. The main contributions of this study are as follows: 1) a novel deep learning architecture by integrating a binary classification network and a ranking network is proposed, superior to existing traditional machine learning algorithms and other state-of-the-art deep learning architectures for biological sequence analysis; 2) the proposed model for blood-secretory protein prediction uses only amino acid sequences, overcoming the heavy dependence of existing methods on annotated protein features; 3) the blood-secretory proteins predicted by our model are statistically significant compared with existing blood-based biomarkers of cancer.

Quantifying Axial Spine Images Using Object-Specific Bi-Path Network.

Automatic estimation of indices from medical images is the main goal of computer-aided quantification (CADq), which speeds up diagnosis and lightens the workload of radiologists. Deep learning technique is a good choice for implementing CADq. Usually, to acquire high-accuracy quantification, specific network architecture needs to be designed for a given CADq task. In this study, considering that the target organs are the intervertebral disc and the dural sac, we propose an object-specific bi-path network (OSBP-Net) for axial spine image quantification. Each path of the OSBP-Net comprises a shallow feature extraction layer (SFE) and a deep feature extraction sub-network (DFE). The SFEs use different convolution strides because the two target organs have different anatomical sizes. The DFEs use average pooling for downsampling based on the observation that the target organs have lower intensity than the background. In addition, an inter-path dissimilarity constraint is proposed and applied to the output of the SFEs, taking into account that the activated regions in the feature maps of two paths should be different theoretically. An inter-index correlation regularization is introduced and applied to the output of the DFEs based on the observation that the diameter and area of the same object express an approximately linear relation. The prediction results of OSBP-Net are compared to several state-of-the-art machine learning-based CADq methods. The comparison reveals that the proposed methods precede other competing methods extensively, indicating its great potential for spine CADq.

Intrinsic brain network alterations in non-clinical adults with a history of childhood trauma

ABSTRACT Background: Childhood trauma is a major social public-health problem worldwide. Previous literature suggests childhood trauma is associated with the development of psychiatric disorders and maladaptive behaviours later in life, but little is known about the neural basis underlying these associations. Objective: The aim of the current study was to investigate intrinsic brain network alterations in non-clinical adults with childhood trauma. Methods: Resting-state functional magnetic resonance imaging (fMRI) data were collected from 65 non-clinical adults with moderate or severe childhood trauma (CT group), according to the international demarcation criteria of the Childhood Trauma Questionnaire (CTQ), and 73 socio-demographically matched non-clinical controls without childhood trauma (no-CT group). Independent component analysis (ICA) was used to extract subnetworks of the default mode network (DMN), salience network (SN), and central executive network (CEN). Results: ICA revealed that the CT group had increased FC of the left medial prefrontal cortex (mPFC) in the anterior DMN (aDMN), increased functional connectivity (FC) of the left anterior insula in the SN, and decreased FC of the inferior parietal gyrus of the right CEN (rCEN). Compared to the controls, the CT group had decreased inter-network FCs between the SN and posterior DMN (pDMN), as well as between the pDMN and rCEN. Conclusions: Impaired FC within the three key brain networks, decreased inter-FC between SN and rCEN, and decreased inter-FC between pDMN and rCEN may reflect biomarkers of childhood trauma.

SiamCAN: Real-Time Visual Tracking Based on Siamese Center-Aware Network.

In this article, we present a novel Siamese center-aware network (SiamCAN) for visual tracking, which consists of the Siamese feature extraction subnetwork, followed by the classification, regression, and localization branches in parallel. The classification branch is used to distinguish the target from background, and the regression branch is introduced to regress the bounding box of the target. To reduce the impact of manually designed anchor boxes to adapt to different target motion patterns, we design the localization branch to localize the target center directly to assist the regression branch generating accurate results. Meanwhile, we introduce the global context module into the localization branch to capture long-range dependencies for more robustness to large displacements of the target. A multi-scale learnable attention module is used to guide these three branches to exploit discriminative features for better performance. Extensive experiments on 9 challenging benchmarks, namely VOT2016, VOT2018, VOT2019, OTB100, LTB35, LaSOT, TC128, UAV123 and VisDrone-SOT2019 demonstrate that SiamCAN achieves leading accuracy with high efficiency. Our source code is available at https://isrc.iscas.ac.cn/gitlab/research/siamcan.

Coupled Network for Robust Pedestrian Detection With Gated Multi-Layer Feature Extraction and Deformable Occlusion Handling.

Pedestrian detection methods have been significantly improved with the development of deep convolutional neural networks. Nevertheless, detecting ismall-scaled pedestrians and occluded pedestrians remains a challenging problem. In this paper, we propose a pedestrian detection method with a couple-network to simultaneously address these two issues. One of the sub-networks, the gated multi-layer feature extraction sub-network, aims to adaptively generate discriminative features for pedestrian candidates in order to robustly detect pedestrians with large variations on scale. The second sub-network targets on handling the occlusion problem of pedestrian detection by using deformable regional region of interest (RoI)-pooling. We investigate two different gate units for the gated sub-network, namely, the channel-wise gate unit and the spatio-wise gate unit, which can enhance the representation ability of the regional convolutional features among the channel dimensions or across the spatial domain, repetitively. Ablation studies have validated the effectiveness of both the proposed gated multi-layer feature extraction sub-network and the deformable occlusion handling sub-network. With the coupled framework, our proposed pedestrian detector achieves promising results on both two pedestrian datasets, especially on detecting small or occluded pedestrians. On the CityPersons dataset, the proposed detector achieves the lowest missing rates (i.e. 40.78% and 34.60%) on detecting small and occluded pedestrians, surpassing the second best comparison method by 6.0% and 5.87%, respectively.

DIGGER: exploring the functional role of alternative splicing in protein interactions.

Alternative splicing plays a major role in regulating the functional repertoire of the proteome. However, isoform-specific effects to protein-protein interactions (PPIs) are usually overlooked, making it impossible to judge the functional role of individual exons on a systems biology level. We overcome this barrier by integrating protein-protein interactions, domain-domain interactions and residue-level interactions information to lift exon expression analysis to a network level. Our user-friendly database DIGGER is available at https://exbio.wzw.tum.de/digger and allows users to seamlessly switch between isoform and exon-centric views of the interactome and to extract sub-networks of relevant isoforms, making it an essential resource for studying mechanistic consequences of alternative splicing.

Image Blind Denoising Using a Generative Adversarial Network for LED Chip Visual Localization

The noise in an LED chip image directly affects its visual positioning accuracy and efficiency. The existing denoising method can improve the image visual quality after removing the noise, but the improvement of the recognition rate and the positioning accuracy of the chip is not apparent. To solve this problem, this paper proposes a convolutional neural network combined with generative adversarial networks framework, the (CNN-GAN-GAN)-based blind denoiser, comprising three subnetworks. First, a noise extraction subnetwork with the proposed adaptive residual dense block extracts noise information from noisy images. Then, the noise modeling subnetwork with a dual-discriminator learns the noise distribution and constructs a paired training dataset. Finally, the blind denoising subnetwork with a hybrid loss function is trained for denoising. Contrast experiments show the proposed method can effectively suppress image noise and artifacts and meet the requirements of LED chip visual positioning accuracy while ensuring chip recognition rate.

JSNet: A simulation network of JPEG lossy compression and restoration for robust image watermarking against JPEG attack

Deep learning-based watermarking methods have achieved a better performance in capacity and invisibility than some traditional methods. However, their robustness against JPEG lossy compression attack is still to be improved. To enhance the robustness and construct an end-to-end method, it is urgent to simulate the JPEG lossy compression by a neural network and then introduce it into the deep learning-based watermarking methods. In this paper, a JPEG simulation network JSNet is proposed to reappear the whole procedure of the JPEG lossy compression and restoration except entropy encoding as realistically as possible. The steps of sampling, DCT, and quantization are modeled by the max-pooling layer, convolution layer, and 3D noise-mask, respectively. The proposed JSNet can simulate JPEG lossy compression with any quality factors. To verify the proposed JSNet in improving the robustness against JPEG compression attack, a CNN-based robust watermarking network (CRWNet) is proposed as an application example. The end-to-end CRWNet contains three subnetworks, i.e., embedding subnetwork, JSNet, and extraction subnetwork. Here. the JSNet is regarded as an attack module in the pipeline. Experimental results on two publicly available datasets (ImageNet, BossBase) demonstrate that: (a) the proposed JSNet can well simulate JPEG lossy compression under any quality factors with small Root mean square error (RMSE) values; (b) the proposed CRWNet considering JSNet has achieved an average 30.6 percent advantage over the basic model without consideration of JSNet.

GEMtractor: extracting views into genome-scale metabolic models.

Computational metabolic models typically encode for graphs of species, reactions and enzymes. Comparing genome-scale models through topological analysis of multipartite graphs is challenging. However, in many practical cases it is not necessary to compare the full networks. The GEMtractor is a web-based tool to trim models encoded in SBML. It can be used to extract subnetworks, for example focusing on reaction- and enzyme-centric views into the model. The GEMtractor is licensed under the terms of GPLv3 and developed at github.com/binfalse/GEMtractor-a public version is available at sbi.uni-rostock.de/gemtractor.

Orientation guided anchoring for geospatial object detection from remote sensing imagery

Object detection from remote sensing imagery plays a significant role in a wide range of applications, including urban planning, intelligent transportation systems, ecology and environment analysis, etc. However, scale variations, orientation variations, illumination changes, and partial occlusions, as well as image qualities, bring great challenges for accurate geospatial object detection. In this paper, we propose an efficient orientation guided anchoring based geospatial object detection network based on convolutional neural networks. To handle objects of varying sizes, the feature extraction subnetwork extracts a pyramid of semantically strong features at different scales. Based on orientation guided anchoring, the anchor generation subnetwork generates a small set of high-quality, oriented anchors as object proposals. After orientation region of interest pooling, objects of interest are detected from the object proposals through the object detection subnetwork. The proposed method has been tested on a large geospatial object detection dataset. Quantitative evaluations show that an overall completeness, correctness, quality, and F1-measure of 0.9232, 0.9648, 0.8931, and 0.9435, respectively, are obtained. In addition, the proposed method achieves a processing speed of 8 images per second on a GPU on the cloud computing platform. Comparative studies with the existing object detection methods also demonstrate the advantageous detection accuracy and computational efficiency of our proposed method.

Analysis of brain subnetworks within the context of their whole-brain networks.

Analyzing the structure and function of the brain from a network perspective has increased considerably over the past two decades, with regional subnetwork analyses becoming prominent in the recent literature. However, despite the fact that the brain, as a complex system of interacting subsystems (i.e., subnetworks), cannot be fully understood by analyzing its constituent parts as independent elements, most studies extract subnetworks from the whole and treat them as independent networks. This approach entails neglecting their interactions with other brain regions and precludes identifying potential compensatory mechanisms outside the analyzed subnetwork. In this study, using simulated and empirical data, we show that the analysis of brain subnetworks within the context of their whole-brain networks, that is, including their interactions with other brain regions, can yield different outcomes when compared to analyzing them as independent networks. We also provide a multivariate mixed-effects modeling framework that allows analyzing subnetworks within the context of their whole-brain networks, and show that it can better disentangle global (whole-brain) and local (subnetwork) differences when compared to standard t-test analyses. T-test analyses may produce misleading results in identifying complex global and local level differences. The provided multivariate model is an extension of a previously developed model for global, system-level hypotheses about the brain. The modified version detailed here provides the same utilities as the original model-quantifying the relationship between phenotypes and brain connectivity, comparing brain networks among groups, predicting brain connectivity from phenotypes, and simulating brain networks-but for local, subnetwork-level hypotheses.

Hierarchical residual learning for image denoising

In recent years, residual learning based convolutional neural networks have been applied to image restoration and achieved some success. To avoid network degradation, deep layers in these methods are identity mappings, which are not easy to be learned as observed in recent image recognition work. In this paper, we propose a novel residual learning based CNN framework for image denoising, which does not need to learn identify mappings while avoiding network degradation. The proposed CNN network contains three kinds of sub-networks: feature extraction sub-network, inference sub-network and fusion sub-network. The feature extraction sub-network is first used to densely extract patches and represent them as high dimensional feature maps. Multiple inference sub-networks are then cascaded to learn noise maps by exploiting multi-scale information in a hierarchical fashion, which makes our method have a strong ability of toleraing errors in noise estimation. Finally, the fusion sub-network fuses the noise maps to obtain the final noise estimation. The proposed hierarchical residual learning network can tackle with multiple general image denoising tasks such as Gaussian denoising and single image super-resolution. Experimental results on several datasets show that our hierarchical residual learning based image denoising method outperforms many state-of-the-art ones.

Altered social cognition and connectivity of default mode networks in the co-occurrence of autistic spectrum disorder and attention deficit hyperactivity disorder.

As two common neurodevelopmental disorders, autistic spectrum disorder and attention deficit hyperactivity disorder frequently occur together. Until now, only a few studies have investigated the co-occurrence of attention deficit hyperactivity disorder and autistic spectrum disorder, this is due to restrictions associated with previous Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition, Text Revision. Most previous research has focused on the developmental trajectories for autistic spectrum disorder and attention deficit hyperactivity disorder separately, while the neural mechanisms underpinning the co-occurrence of autistic spectrum disorder and attention deficit hyperactivity disorder remain largely unknown. We studied 162 autistic spectrum disorder individuals (including 79 co-attention deficit hyperactivity disorder and 83 non-attention deficit hyperactivity disorder patients) and 177 typical developing individuals using resting-state functional magnetic resonance imaging data from the Autism Brain Imaging Data Exchange II, an aggregated magnetic resonance imaging dataset from 19 centers. Independent component analysis was used to extract sub-networks from the classic resting-state networks. Functional connectivity values within (intra-iFC) and between (inter-iFC) these networks were then determined. Subsequently, we compared the ASD_coADHD group with the ASD_nonADHD group in relation to the abnormal intra-iFC and inter-iFC of autistic spectrum disorder group relative to the typical developing group. The ASD_coADHD group showed more severe social impairment and decreased intra-iFC in the bilateral posterior cingulate cortex of the default mode network (independent component 17) and increased inter-iFC between the default mode network (independent component 8) and the somatomotor networks (independent component 2) compared to the ASD_nonADHD group. In addition, the strength of the intra-iFC in the default mode network was associated with the severity of autistic traits across the entire autistic spectrum disorder group and particularly the ASD_coADHD group. Our results showed that dysfunction of the default mode network is a central feature in the co-occurrence of autistic spectrum disorder and attention deficit hyperactivity disorder, including connectivity within the default mode network as well as between the default mode network and the somatomotor networks, thus supporting the existence of a clinically combined phenotype (autistic spectrum disorder + attention deficit hyperactivity disorder).