CNN-based Network Research Articles

LSLPCT: An Enhanced Local Semantic Learning Transformer for 3-D Point Cloud Analysis

The 3D point cloud is a common 3D data representation that has received increasing attention for remote sensing applications. However, processing 3D point cloud semantics, especially local semantic information, has always been a challenge and has attracted much attention. In this paper, we propose a novel enhanced local semantic learning transformer for 3D point cloud analysis, which aims to enhance the transformer awareness of local semantic features to handle complex point cloud tasks. First, we propose a novel transformer framework, the local semantic learning point cloud transformer (LSLPCT), which not only learns 3D point clouds the global information, but also enhances the perception of local semantic information end-to-end. Second, we design an efficient local semantic learning self-attention mechanism, namely LSL-SA, which can parallelize the perception of global contextual information and the capture of finer-grained local semantic features. Third, our proposed LSL-SA is easy to implement and can integrate existing transformers and CNN-based networks for processing various point cloud tasks. Numerous experiments in different types of point cloud tasks have been conducted, and our method performs better or is competitive with other state-of-the-art methods.

CNN-Based Network Application for Petrophysical Parameter Inversion: Sensitivity Analysis of Input–Output Parameters and Network Architecture

Accurate estimation of petrophysical properties (e.g., porosity, clay volume) of subsurface rock from seismic data/elastic properties is significant to reservoir characterization. Conventional model-driven inversion strategies for estimating petrophysical parameters confront with the deficiency of prior knowledge. In contrast, machine learning-based approaches are adapted to account for reservoir parameter estimation through developing nonlinear mapping and quantifying uncertainty. However, most of the current researches mainly concentrates on the single parameter prediction with different neural network architectures, which, in turn, conflicts with the truth of coupling multiple reservoir properties. To quantify the sensitivity of input-output parameters and the effects of network architecture on the accuracy of petrophysical parameter inversion, we propose a CNN-based network strategy to estimate multiple reservoir parameters simultaneously. The results from both synthetic labeled data and field data and uncertainty analysis strongly demonstrate that SopenCNN, abbreviated from multiple input and single output openCNN, exhibits the highest prediction accuracy, while the cycleCNN with multiple input and multiple output, referred to as McycleCNN, is superior to the MopenCNN, which means that an openCNN contains multiple input and multiple output. It means that, for similar network architecture, the number of input and output parameters makes a significant impact on prediction accuracy. Moreover, for similar multiple inputs and outputs, the fine-tuning McycleCNN, parallelly updated in each intermediate closed-loop step, behaves much better accordingly. The application of the three workflows with varying architecture on field tight sandstone reservoirs demonstrates that network based inversion strategy could establish a mapping function to characterize spatially varying reservoir parameters.

Attention-based DenseNet network for multi-source remote sensing classification

Remote sensing data are abundant in spatial, image information and widely applied in earth observation covering military and civil fields. As the hot spot and critical point, scene classification has attracted increasing attention. Convolutional neural networks (CNNs) behave as representative technologies in scene classification. Nevertheless, many CNNs are lack of the ability to distinguish key information from redundant information. To overcome the problem and achieve efficient feature extractions from multi-source data, we integrate DenseNet and attention mechanism into a CNN-based network. Multi-scale convolution and threedimensional convolution are applied to extract features; and then, to explore and obtain feature information in a deeper level, a Dense structure are utilized; Afterwards, the features extracted from previous procedures are weighed due to importance and merged through the convolutional attention model. The experiment validates that the presented method poses advantages of efficiently merging various feature by Dense structure and the convolutional attention model.

Leveraging network using controlled weight learning approach for thyroid cancer lymph node detection

Identification of cervical metastatic lymph nodes (LN) on I-131 post-ablation whole-body planar scans (WBS) for cancer staging is crucial for patients with papillary thyroid cancer (PTC). The existing deep network is plagued by instability in finding the under-represented LN classes on highly complex WBS, where end-to-end tuned models that can account for this uncertainty region for multi-class locations are needed. Hence, as a key contribution of this study, we designed a novel leveraging segmentation network with input guidance (LSIG) end-to-end training model without pre and post-processing features that can learn ideal parameter settings depending on the quantity of multiple-object instances. To improve the co-occurrence of classes and control the false positive regions, we proposed a re-weighting negative control (RNC) mechanism that combines two key components, namely the re-weighting (Rw) term and the negative control function (NcF). This unified approximation of weighted training would leverage the network to control and learn the desired weights of true positives towards the LN region. As an end-to-end network training, we utilize a Unet-like convolution neural network (CNN) model. The performance of the LSIG is compared with the CNN-based networks, based on the ground truth (GT) mask developed using post-ablation single-photon emission computed tomography (SPECT/CT). Furthermore, the effectiveness of the two components used in the LSIG framework is evaluated on WBS datasets. Our proposed LSIG with a fully guided (Fg) LS-FgCNN model yielded a superior performance with high AUC value of 94.9%, which is 14.6% higher than the previous network for PTC.

Lithology identification from well-log curves via neural networks with additional geologic constraint

Lithology identification is of great importance in reservoir characterization. Recently, many researchers have applied machine-learning techniques to solve lithology identification problems from well-log curves, and their works indicate three main characteristics. First, most works predict lithofacies using features measured during logging, whereas very few consider adding stratigraphic sequence information that is available prior to drilling to solve this problem. Second, most studies predict lithofacies using measured properties of one depth point, whereas few take the influence of the neighboring formation into account. Third, due to a lack of publicly available interpreted well-log data, previous research has concentrated on applying different algorithms on their private data set, making it impossible to perform a comparison. We have developed a machine-learning framework to solve the lithology classification problem from well-log curves by incorporating an additional geologic constraint. The constraint is a stratigraphic unit, and we use it as an additional feature. We evaluate three types of recurrent neural networks (RNNs), bidirectional long short-term memory, bidirectional gated recurrent unit (Bi-GRU), and GRU-based encoder-decoder architecture with attention, as well as two types of 1D convolutional neural networks (1D CNNs), temporal convolutional network and multiscale residual network, on a publicly available data set from the North Sea. The RNN-based networks and 1D CNN-based networks can process sequential data, enabling the model to have access to information from neighboring formations when performing lithofacies prediction at a particular depth. Our experiments indicate that geologic constraint improves the performance of the models significantly, and that the overall performance of RNN-based networks is better and more consistent.

Automated detection and analysis of social behaviors among preweaning piglets using key point-based spatial and temporal features

In the pig industry, the social behaviors of preweaning piglets are critical indicators of their liveability, growth, health, and welfare status. In this study, we developed a novel method based on convolutional neural networks (CNNs) that extracts high-quality spatiotemporal features to detect different types of social behaviors among preweaning piglets: snout-snout and snout-body social nosing, and snout-snout and snout-body aggressive/playing behavior. In our method, piglets were first detected and individually tracked using an integrated CNN-based network. Upon piglet detection, the key points of the piglet body were detected using a key point detector and a 1-second-tracking-unit strategy was adopted to reduce the influence of tracking errors and missing the detection of piglets. To detect suspected social behaviors in a tracking unit, we constructed a self-adaptive spatial affinity kernel function to represent the relationships between piglets, and pairwise piglets with spatial affinity higher than 0.7 were kept for further analysis, while the others were filtered out. Thereafter, high-quality spatiotemporal features consisting of the piglets’ spatial affinity and the corresponding linear motion and angular motion intensities were extracted and input into a support vector machine to hierarchically detect and classify piglet social behaviors. In addition, the proportions of the four detected social behaviors associated with the interaction points were analyzed in short video clips and a long video that covered the whole daytime. Our method achieved a favorable detection performance with a recall of 0.9687, a precision of 0.9309, and an F1 of 0.9494 in the short video clips, and a recall of 0.9576, a precision of 0.9187, and an F1 of 0.9377 in the 8-h long video episode. This shows the feasibility of the detection of social behaviors for piglets using CNNs.

Sparsity-aware channel estimation for mmWave massive MIMO: A deep CNN-based approach

The deep convolutional neural network (CNN) is exploited in this work to conduct the challenging channel estimation for mmWave massive multiple input multiple output (MIMO) systems. The inherent sparse features of the mmWave massive MIMO channels can be extracted and the sparse channel supports can be learnt by the multi-layer CNN-based network through training. Then accurate channel inference can be efficiently implemented using the trained network. The estimation accuracy and spectrum efficiency can be further improved by fully utilizing the spatial correlation among the sparse channel supports of different antennas. It is verified by simulation results that the proposed deep CNN-based scheme significantly outperforms the state-of-the-art benchmarks in both accuracy and spectrum efficiency.

Efficient residual attention network for single image super-resolution

The use of deep convolutional neural networks (CNNs) for image super-resolution (SR) from low-resolution (LR) input has achieved remarkable reconstruction performance with the utilization of residual structures and visual attention mechanisms. However, existing single image super-resolution (SISR) methods with deeper network architectures can encounter representational bottlenecks in CNN-based networks and neglect model efficiency in model statistical inference. To solve these issues, in this paper, we design a channel hourglass residual structure (CHRS) and explore an efficient channel attention (ECA) mechanism to extract more representative features and ease the computational burden. Specifically, our CHRS, consisting of several nested residual modules, is developed to learn more discriminative representations with fewer model parameters, and the ECA is presented to efficiently capture local cross-channel interaction by subtly applying 1D convolution. Finally, we propose an efficient residual attention network (ERAN), which not only fully learns more representative features but also pays special attention to network learning efficiency. Extensive experiments demonstrate that our ERAN achieves certain improvements in model performance and implementation efficiency compared to other previous state-of-the-art methods.

Deep multi-feature learning architecture for water body segmentation from satellite images

Automatic water body extraction from satellite images of various scenes is a classical and challenging task in remote sensing and image interpretation. Convolutional neural network (CNN) has become prominent option for performing image segmentation task in remote sensing applications. However, CNN-based networks have non-trivial issues for segmenting such as: (1) blurring boundary pixels; (2) large number of trainable parameters; and (3) huge number of training samples. In this paper, we propose an end-to-end multi-feature based CNN architecture, called as W-Net, to perform water body segmentation. W-Net consists of contracting/expanding networks and inception layers. W-Net takes advantage of contracting network to capture context information while localization is achieved with expanding network. With these networks, W-Net is able to train on less number of images and extract water pixels accurately. Use of inception layers reduces computational burden within the network by decreasing total number of trainable parameters. W-Net incorporated two refinement modules to enhance predicted results which mitigate blurring effect and to inspect continuity of boundary pixels. Dataset consisting 2671 images with manually annotated ground truths are built to validate performance and effectiveness of our proposed method. In addition, we evaluated our method on crack detection dataset where W-Net achieved competitive performance with Deepcrack. W-Net accomplished excellent performance on the water body dataset (I∕U=0.9434 and F−score=0.9509).

Weakly Supervised Object Detection Using Proposal- and Semantic-Level Relationships.

In recent years, weakly supervised object detection has attracted great attention in the computer vision community. Although numerous deep learning-based approaches have been proposed in the past few years, such an ill-posed problem is still challenging and the learning performance is still behind the expectation. In fact, most of the existing approaches only consider the visual appearance of each proposal region but ignore to make use of the helpful context information. To this end, this paper introduces two levels of context into the weakly supervised learning framework. The first one is the proposal-level context, i.e., the relationship of the spatially adjacent proposals. The second one is the semantic-level context, i.e., the relationship of the co-occurring object categories. Therefore, the proposed weakly supervised learning framework contains not only the cognition process on the visual appearance but also the reasoning process on the proposal- and semantic-level relationships, which leads to the novel deep multiple instance reasoning framework. Specifically, built upon a conventional CNN-based network architecture, the proposed framework is equipped with two additional graph convolutional network-based reasoning models to implement object location reasoning and multi-label reasoning within an end-to-end network training procedure. Comprehensive experiments on the widely used PASCAL VOC and MS COCO benchmarks have been implemented, which demonstrate the superior capacity of the proposed approach when compared with other state-of-the-art methods and baseline models.

Retracted: An easy-to-use deep-learning model for highly accurate diagnosis of Parkinson's disease using SPECT images

Accurate diagnosis of Parkinson's Disease (PD) at its early stages remains a challenge for modern clinicians. In this study, we utilize a convolutional neural network (CNN) approach to address this problem. In particular, we develop a CNN-based network model highly capable of discriminating PD patients based on Single Photon Emission Computed Tomography (SPECT) images from healthy controls. A total of 2723 SPECT images are analyzed in this study, of which 1364 images from the healthy control group, and the other 1359 images are in the PD group. Image normalization process is carried out to enhance the regions of interests (ROIs) necessary for our network to learn distinguishing features from them. A 10-fold cross-validation is implemented to evaluate the performance of the network model. Our approach demonstrates outstanding performance with an accuracy of 99.34 %, sensitivity of 99.04 % and specificity of 99.63 %, outperforming all previously published results. Given the high performance and easy-to-use features of our network, it can be deduced that our approach has the potential to revolutionize the diagnosis of PD and its management.

Ionospheric correction in P‐band ISAR imaging based on polar formatting algorithm and convolutional neural network

The ionosphere causes serious phase error in P-band inverse synthetic aperture radar (ISAR) systems, which makes it difficult to obtain a high-quality image. Recently, the convolutional neural network (CNN) has gained much attention in signal processing, and it can automatically extract features to realise an image super-resolution reconstruction. As a popular CNN-based network, U-net can work with less training samples. Hence, the authors are interested in exploiting and modifying the U-net to enhance the P-band ISAR imaging. In this study, in light of the analysis of the effect of the ionospheric total electron content on the ground-based P-band radar echo signal, a novel ISAR imaging method is proposed for the ionospheric effect correction based on the modified U-net and polar formatting algorithm (PFA). The PFA is performed for the phase error coarse compensation. Then, the phase error fine compensation is exploited by the trained U-net. The proposed method can adapt the ionosphere disturbances and show high performance in imaging quality and computational efficiency. The simulation results show the effectiveness of the proposed method.

CNN-Based Network Intrusion Detection against Denial-of-Service Attacks

As cyberattacks become more intelligent, it is challenging to detect advanced attacks in a variety of fields including industry, national defense, and healthcare. Traditional intrusion detection systems are no longer enough to detect these advanced attacks with unexpected patterns. Attackers bypass known signatures and pretend to be normal users. Deep learning is an alternative to solving these issues. Deep Learning (DL)-based intrusion detection does not require a lot of attack signatures or the list of normal behaviors to generate detection rules. DL defines intrusion features by itself through training empirical data. We develop a DL-based intrusion model especially focusing on denial of service (DoS) attacks. For the intrusion dataset, we use KDD CUP 1999 dataset (KDD), the most widely used dataset for the evaluation of intrusion detection systems (IDS). KDD consists of four types of attack categories, such as DoS, user to root (U2R), remote to local (R2L), and probing. Numerous KDD studies have been employing machine learning and classifying the dataset into the four categories or into two categories such as attack and benign. Rather than focusing on the broad categories, we focus on various attacks belonging to same category. Unlike other categories of KDD, the DoS category has enough samples for training each attack. In addition to KDD, we use CSE-CIC-IDS2018 which is the most up-to-date IDS dataset. CSE-CIC-IDS2018 consists of more advanced DoS attacks than that of KDD. In this work, we focus on the DoS category of both datasets and develop a DL model for DoS detection. We develop our model based on a Convolutional Neural Network (CNN) and evaluate its performance through comparison with an Recurrent Neural Network (RNN). Furthermore, we suggest the optimal CNN design for the better performance through numerous experiments.

Lung Nodule Detection from Feature Engineering to Deep Learning in Thoracic CT Images: a Comprehensive Review.

This paper presents a systematic review of the literature focused on the lung nodule detection in chest computed tomography (CT) images. Manual detection of lung nodules by the radiologist is a sequential and time-consuming process. The detection is subjective and depends on the radiologist's experiences. Owing to the variation in shapes and appearances of a lung nodule, it is very difficult to identify the proper location of the nodule from a huge number of slices generated by the CT scanner. Small nodules (< 10mm in diameter) may be missed by this manual detection process. Therefore, computer-aided diagnosis (CAD) system acts as a "second opinion" for the radiologists, by making final decision quickly with higher accuracy and greater confidence. The goal of this survey work is to present the current state of the artworks and their progress towards lung nodule detection to the researchers and readers in this domain. This review paper has covered the published works from 2009 to April 2018. Different nodule detection approaches are described elaborately in this work. Recently, it is observed that deep learning (DL)-based approaches are applied extensively for nodule detection and characterization. Therefore, emphasis has been given to convolutional neural network (CNN)-based DL approaches by describing different CNN-based networks.

4-D SLAM: An Efficient Dynamic Bayes Network-Based Approach for Dynamic Scene Understanding

Most of the existing simultaneous localization and mapping (SLAM) methods are based on the static environment assumption. The presence of moving objects in the scene will lead to much uncertainty in SLAM results, which also hinders the loop-closure detection (LCD). Although moving object tracking (MOT) is necessary for planning and decisions, it is often accomplished separately. To jointly solve SLAM and MOT for complex urban driving scenarios, this paper presents a high performance method named 4D-SLAM based on the fusion of LiDAR and IMU. The integration of SLAM and MOT is formulated as a joint posterior probability problem based on a dynamic Bayesian network (DBN) and is implemented with the following four sequential stages: preprocessing, moving object detection and tracking, odometry estimation and mapping. In the preprocessing stage, the motion distortion uncertainty caused by LiDAR scanning is first compensated for, and the initial LiDAR motion is estimated. In the moving object detection and tracking stage, we exploit a CNN-based segmentation network to detect the potential moving objects first, then the states of the potential moving objects are optimized by an unscented Kalman filter (UKF). In the odometry estimation stage, the distinctive planar and edge features extracted from a static background point cloud are used for the odometry estimation, and a two-step Levenberg-Marquardt optimization method is adopted to solve the 6DOF pose across consecutive scans. In the mapping stage, the mapping based on the pose estimation result and LCD are realized, and graph-based global optimization is exploited to further improve the map consistency for large scale environment. The comprehensive experiments with the open source dataset KITTI and the data collected by us show that the presented method not only outperforms the SOTA SLAM methods in terms of trajectory and mapping accuracy but can also detect and track moving objects efficiently.

Aspect-Based Sentiment Classification Using Interactive Gated Convolutional Network

Aspect-based sentiment classification aims to detect the sentiment polarity of a target in a given context. Most previous approaches use long short-term memory (LSTM) and attention mechanisms to predict the sentiment polarity of targets, which are usually complex and need more training time. Some previous approaches are based on convolutional neural networks (CNN) and gating mechanisms, which are much simpler, efficient and takes lesser convergence time than LSTM due to parallelized computations during training. However, such CNN-based networks ignore the separate modeling of targets via context-specific representations. In this paper, we propose a novel interactive gated convolutional network (IGCN) that uses a bidirectional gating mechanism to learn mutual relation between the target and corresponding review context. IGCN also uses positional information of context words with respect to the given target, POS tags, and domain-specific word embeddings for predicting the sentiment of a target. The experimental results on SemEval 2014 datasets show the effectiveness of our proposed IGCN model.

Cascaded Dense-UNet for Image Super-Resolution

Recently, deep convolutional neural networks (CNNs) have achieved great success in single image super-resolution (SISR). Especially, dense skip connections and residual learning structures promote better performance. While most existing deep CNN-based networks exploit the interpolation of upsampled original images, or do transposed convolution in the reconstruction stage, which do not fully employ the hierarchical features of the networks for final reconstruction. In this paper, we present a novel cascaded Dense-UNet (CDU) structure to take full advantage of all hierarchical features for SISR. In each Dense-UNet block (DUB), many short, dense skip pathways can facilitate the flow of information and integrate the different receptive fields. A series of DUBs are concatenated to acquire high-resolution features and capture complementary contextual information. Upsampling operators are in DUBs. Furthermore, residual learning is introduced to our network, which can fuse shallow features from low resolution (LR) image and deep features from cascaded DUBs to further boost super-resolution (SR) reconstruction results. The proposed method is evaluated quantitatively and qualitatively on four benchmark datasets, our network achieves comparable performance to state-of-the-art super-resolution approaches and obtains pleasant visualization results.

Single Space Object Image Denoising and Super-Resolution Reconstructing Using Deep Convolutional Networks

Space object recognition is the basis of space attack and defense confrontation. High-quality space object images are very important for space object recognition. Because of the large number of cosmic rays in the space environment and the inadequacy of optical lenses and detectors on satellites to support high-resolution imaging, most of the images obtained are blurred and contain a lot of cosmic-ray noise. So, denoising methods and super-resolution methods are two effective ways to reconstruct high-quality space object images. However, most super-resolution methods could only reconstruct the lost details of low spatial resolution images, but could not remove noise. On the other hand, most denoising methods especially cosmic-ray denoising methods could not reconstruct high-resolution details. So in this paper, a deep convolutional neural network (CNN)-based single space object image denoising and super-resolution reconstruction method is presented. The noise is removed and the lost details of the low spatial resolution image are well reconstructed based on one very deep CNN-based network, which combines global residual learning and local residual learning. Based on a dataset of satellite images, experimental results demonstrate the feasibility of our proposed method in enhancing the spatial resolution and removing the noise of the space objects images.

GRN: Gated Relation Network to Enhance Convolutional Neural Network for Named Entity Recognition

The dominant approaches for named entity recognitionm (NER) mostly adopt complex recurrent neural networks (RNN), e.g., long-short-term-memory (LSTM). However, RNNs are limited by their recurrent nature in terms of computational efficiency. In contrast, convolutional neural networks (CNN) can fully exploit the GPU parallelism with their feedforward architectures. However, little attention has been paid to performing NER with CNNs, mainly owing to their difficulties in capturing the long-term context information in a sequence. In this paper, we propose a simple but effective CNN-based network for NER, i.e., gated relation network (GRN), which is more capable than common CNNs in capturing long-term context. Specifically, in GRN we firstly employ CNNs to explore the local context features of each word. Then we model the relations between words and use them as gates to fuse local context features into global ones for predicting labels. Without using recurrent layers that process a sentence in a sequential manner, our GRN allows computations to be performed in parallel across the entire sentence. Experiments on two benchmark NER datasets (i.e., CoNLL2003 and Ontonotes 5.0) show that, our proposed GRN can achieve state-of-the-art performance with or without external knowledge. It also enjoys lower time costs to train and test.

Vibration events recognition of optical fiber based on multi-scale 1-D CNN

A new CNN-based deep neural network, multi-scale one-dimensional convolutional neural network (MS 1-D CNN) was proposed to improve the efficiency and accuracy of vibration event recognition for a phase-sensitive optical time-domain reflectometry (Φ-OTDR) distributed optical fiber vibration sensing system. The raw vibration signals are pre-processed first to remove noise as far as possible. The pre-processing operations include pre-emphasis filtering, normalization and spectral subtraction. The pre-processed signals are used as the inputs of MS 1-D CNN directly. MS 1-D CNN realizes the end-to-end feature extraction of vibration signals and finally recognizes the vibration events by using a fully-connected layer (FC layer) and a Softmax layer. In comparison with two-dimensional convolutional neural network (2-D CNN) and one-dimensional convolutional neural network (1-D CNN), the proposed method balances the time and frequency scales better during feature extraction and reduces the pending parameters of the whole neural network. A vibration recognition experiment was designed to classify the three types of the vibration events including damaging, knocking and interference. The recognition results show that MS 1-D CNN achieves similar recognition performance, over 96 percent, at twice processing speed compared to 2-D CNN. Therefore, it is beneficial to improve the real-timing of vibration monitoring while maintaining the recognition performance.