Generalization Capability Of Model Research Articles

A Residual Attention and Local Context-Aware Network for Road Extraction from High-Resolution Remote Sensing Imagery

Extracting road information from high-resolution remote sensing images (HRI) can provide crucial geographic information for many applications. With the improvement of remote sensing image resolution, the image data contain more abundant feature information. However, this phenomenon also enhances the spatial heterogeneity between different types of roads, making it difficult to accurately discern the road and non-road regions using only spectral characteristics. To remedy the above issues, a novel residual attention and local context-aware network (RALC-Net) is proposed for extracting a complete and continuous road network from HRI. RALC-Net utilizes a dual-encoder structure to improve the feature extraction capability of the network, whose two different branches take different feature information as input data. Specifically, we construct the residual attention module using the residual connection that can integrate spatial context information and the attention mechanism, highlighting local semantics to extract local feature information of roads. The residual attention module combines the characteristics of both the residual connection and the attention mechanism to retain complete road edge information, highlight essential semantics, and enhance the generalization capability of the network model. In addition, the multi-scale dilated convolution module is used to extract multi-scale spatial receptive fields to improve the model’s performance further. We perform experiments to verify the performance of each component of RALC-Net through the ablation study. By combining low-level features with high-level semantics, we extract road information and make comparisons with other state-of-the-art models. The experimental results show that the proposed RALC-Net has excellent feature representation ability and robust generalizability, and can extract complete road information from a complex environment.

Lossless Coding of Point Cloud Geometry Using a Deep Generative Model

This paper proposes a lossless point cloud (PC) geometry compression method that uses neural networks to estimate the probability distribution of voxel occupancy. First, to take into account the PC sparsity, our method adaptively partitions a point cloud into multiple voxel block sizes. This partitioning is signalled via an octree. Second, we employ a deep auto-regressive generative model to estimate the occupancy probability of each voxel given the previously encoded ones. We then employ the estimated probabilities to code efficiently a block using a context-based arithmetic coder. Our context has variable size and can expand beyond the current block to learn more accurate probabilities. We also consider using data augmentation techniques to increase the generalization capability of the learned probability models, in particular in the presence of noise and lower-density point clouds. Experimental evaluation, performed on a variety of point clouds from four different datasets and with diverse characteristics, demonstrates that our method reduces significantly (by up to 37%) the rate for lossless coding compared to the state-of-the-art MPEG codec.

Super-resolution of near-surface temperature utilizing physical quantities for real-time prediction of urban micrometeorology

The present paper proposes a super-resolution (SR) model based on a convolutional neural network and applies it to the near-surface temperature in urban areas. The SR model incorporates a skip connection, a channel attention mechanism, and separated feature extractors for the inputs of temperature, building height, downward shortwave radiation, and horizontal velocity. We train the SR model with sets of low-resolution (LR) and high-resolution (HR) images from building-resolving large-eddy simulations (LESs) in a city, where the horizontal resolutions of LR and HR are 20 and 5 m, respectively The generalization capability of the SR model is confirmed with LESs in another city. The estimated HR temperature fields are more accurate than those of the bicubic interpolation and image SR model that takes only the temperature as its input. Except for the temperature input, the building height is the most important to reconstruct the HR temperature and enables the SR model to reduce errors in temperature near building boundaries. The SR model considers the appropriate boundary for each building from its height information. The analysis of attention weights indicates that the importance of the building height increases as the downward shortwave radiation becomes larger. The contrast between sun and shade is strengthened with the increase in solar radiation, which may affect the temperature distribution. The short inference time suggests the potential of the proposed SR model to facilitate a real-time HR prediction in metropolitan areas by combining it with an LR building-resolving LES model.

Prediction of House Price Index Based on Bagging Integrated WOA-SVR Model

Aiming at the shortcomings of a single machine learning model with low model prediction accuracy and insufficient generalization ability in house price index prediction, a whale algorithm optimized support vector regression model based on bagging ensemble learning method is proposed. Firstly, gray correlation analysis is used to obtain the main influencing factors of house prices, and the segmentation forecasting method is used to divide the data set and forecast the house prices in the coming year using the data of the past ten years. Secondly, the whale optimization algorithm is used to find the optimal parameters of the penalty factor and kernel function in the SVR model, and then, the WOA-SVR model is established. Finally, in order to further improve the model generalization capability, a bagging integration strategy is used to further integrate and optimize the WOA-SVR model. The experiments are conducted to forecast the house price indices of four regions, Beijing, Shanghai, Tianjin, and Chongqing, respectively, and the results show that the prediction accuracy of the proposed integrated model is better than the comparison model in all cases.

STC-Det: A Slender Target Detector Combining Shadow and Target Information in Optical Satellite Images

Object detection has made great progress. However, due to the unique imaging method of optical satellite remote sensing, the detection of slender targets is still insufficient. Specifically, the perspective of optical satellites is small, and the characteristics of slender targets are severely lost during imaging, resulting in insufficient detection task information; at the same time, the appearance of slender targets in the image is greatly affected by the satellite perspective, which is likely to cause insufficient generalization capabilities of conventional detection models. In response to these two points, we have made some improvements. First, in this paper, we introduce the shadow as auxiliary information to complement the trunk features of the target lost in imaging. Second, to reduce the impact of satellite perspective on imaging, in this paper, we use the characteristic that shadow information is not affected by satellite perspective to design STC-Det. STC-Det treats the shadow and the target as two different types of targets and uses the shadow information to assist the detection, reducing the impact of the satellite perspective on detection. Among them, in order to improve the performance of STC-Det, we propose an automatic matching method (AMM) of shadow and target and a feature fusion method (FFM). Finally, this paper proposes a new method to calculate the heatmaps of detectors, which verifies the effectiveness of the proposed network in a visual way. Experiments show that when the satellite perspective is variable, the precision of STC-Det is increased by 1.7%, and when the satellite perspective is small, the precision of STC-Det is increased by 5.2%.

Discriminative adversarial domain generalization with meta-learning based cross-domain validation

The generalization capability of machine learning models, which refers to generalizing the knowledge for an “unseen” domain via learning from one or multiple seen domain(s), is of great importance to develop and deploy machine learning applications in the real-world conditions. Domain Generalization (DG) techniques aim to enhance such generalization capability of machine learning models, where the learnt feature representation and the classifier are two crucial factors to improve generalization and make decisions. In this paper, we propose Discriminative Adversarial Domain Generalization (DADG) with meta-learning-based cross-domain validation. Our proposed framework tries to learn a domain-invariant feature representation from source domains and generalize it to the unseen domains. It contains two main components that work synergistically to build a domain-generalized Deep Neural Network (DNN) model: (i) discriminative adversarial learning, which proactively learns a generalized feature representation on multiple “seen” domains, and (ii) meta-learning based cross domain validation, which simulates train/test domain shift via applying meta-learning techniques in the training process. In the experimental evaluation, a comprehensive comparison has been made among our proposed approach and other existing approaches on three benchmark datasets. The results shown that DADG consistently outperforms a strong baseline DeepAll, and outperforms the other existing DG algorithms in most of the evaluation cases.

Deep Learning-Based Automated Thrombolysis in Cerebral Infarction Scoring: A Timely Proof-of-Principle Study.

Mechanical thrombectomy is an established procedure for treatment of acute ischemic stroke. Mechanical thrombectomy success is commonly assessed by the Thrombolysis in Cerebral Infarction (TICI) score, assigned by visual inspection of X-ray digital subtraction angiography data. However, expert-based TICI scoring is highly observer-dependent. This represents a major obstacle for mechanical thrombectomy outcome comparison in, for instance, multicentric clinical studies. Focusing on occlusions of the M1 segment of the middle cerebral artery, the present study aimed to develop a deep learning (DL) solution to automated and, therefore, objective TICI scoring, to evaluate the agreement of DL- and expert-based scoring, and to compare corresponding numbers to published scoring variability of clinical experts. The study comprises 2 independent datasets. For DL system training and initial evaluation, an in-house dataset of 491 digital subtraction angiography series and modified TICI scores of 236 patients with M1 occlusions was collected. To test the model generalization capability, an independent external dataset with 95 digital subtraction angiography series was analyzed. Characteristics of the DL system were modeling TICI scoring as ordinal regression, explicit consideration of the temporal image information, integration of physiological knowledge, and modeling of inherent TICI scoring uncertainties. For the in-house dataset, the DL system yields Cohen’s kappa, overall accuracy, and specific agreement values of 0.61, 71%, and 63% to 84%, respectively, compared with the gold standard: the expert rating. Values slightly drop to 0.52/64%/43% to 87% when the model is, without changes, applied to the external dataset. After model updating, they increase to 0.65/74%/60% to 90%. Literature Cohen’s kappa values for expert-based TICI scoring agreement are in the order of 0.6. The agreement of DL- and expert-based modified TICI scores in the range of published interobserver variability of clinical experts highlights the potential of the proposed DL solution to automated TICI scoring.

A Regional Wind Power Probabilistic Forecast Method Based on Deep Quantile Regression

Different from individual wind power forecast and regional wind power forecast (RWPF), one of the most significant research articles for alleviating negative influence on power systems aims to estimate the generation of multiple wind farms in the specific region, which is a valuable complement of the wind power forecast. This article proposes a nonparametric probabilistic method for RWPF, a quantile regression neural network (QRNN), enhancing the abilities of nonlinear mapping and massive data dealing. On this basis, the deep quantile regression is proposed to improve the performance of the QRNN. In this approach, the local-connected method is applied to the input layer of the neural network for tackling the challenge of the massive data. A ramp function is designed to avoid multiple quantile curves crossing problem. To improve the model's generalization capability, a smoothing method is applied to the loss function for achieving differentiability everywhere. By properly constructing the model, the approach provides a specific solution for RWPF with the massive input information. The test results on a region with ten wind farms demonstrate the effectiveness of the proposed approach.

A Contrast Augmentation Approach to Improve Multi-Scanner Generalization in MRI.

Most data-driven methods are very susceptible to data variability. This problem is particularly apparent when applying Deep Learning (DL) to brain Magnetic Resonance Imaging (MRI), where intensities and contrasts vary due to acquisition protocol, scanner- and center-specific factors. Most publicly available brain MRI datasets originate from the same center and are homogeneous in terms of scanner and used protocol. As such, devising robust methods that generalize to multi-scanner and multi-center data is crucial for transferring these techniques into clinical practice. We propose a novel data augmentation approach based on Gaussian Mixture Models (GMM-DA) with the goal of increasing the variability of a given dataset in terms of intensities and contrasts. The approach allows to augment the training dataset such that the variability in the training set compares to what is seen in real world clinical data, while preserving anatomical information. We compare the performance of a state-of-the-art U-Net model trained for segmenting brain structures with and without the addition of GMM-DA. The models are trained and evaluated on single- and multi-scanner datasets. Additionally, we verify the consistency of test-retest results on same-patient images (same and different scanners). Finally, we investigate how the presence of bias field influences the performance of a model trained with GMM-DA. We found that the addition of the GMM-DA improves the generalization capability of the DL model to other scanners not present in the training data, even when the train set is already multi-scanner. Besides, the consistency between same-patient segmentation predictions is improved, both for same-scanner and different-scanner repetitions. We conclude that GMM-DA could increase the transferability of DL models into clinical scenarios.

A CNN-Based System for Mobile Robot Navigation in Indoor Environments via Visual Localization with a Small Dataset

Deep learning has made great advances in the field of image processing, which allows automotive devices to be more widely used in humans’ daily lives than ever before. Nowadays, the mobile robot navigation system is among the hottest topics that researchers are trying to develop by adopting deep learning methods. In this paper, we present a system that allows the mobile robot to localize and navigate autonomously in the accessible areas of an indoor environment. The proposed system exploits the Convolutional Neural Network (CNN) model’s advantage to extract data feature maps for image classification and visual localization, which attempts to precisely determine the location region of the mobile robot focusing on the topological maps of the real environment. The system attempts to precisely determine the location region of the mobile robot by integrating the CNN model and topological map of the robot workspace. A dataset with small numbers of images is acquired from the MYNT EYE camera. Furthermore, we introduce a new loss function to tackle the bounded generalization capability of the CNN model in small datasets. The proposed loss function not only considers the probability of the input data when it is allocated to its true class but also considers the probability of allocating the input data to other classes rather than its actual class. We investigate the capability of the proposed system by evaluating the empirical studies based on provided datasets. The results illustrate that the proposed system outperforms other state-of-the-art techniques in terms of accuracy and generalization capability.

TGAN: A simple model update strategy for visual tracking via template-guidance attention network

Visual attention has been widely used in various fields of visual tasks in recent years. Recently, visual trackers based on probabilistic discriminative model prediction (PrDiMP) and Siamese box adaptive network (SiamBAN) have attracted much attention due to their excellent performance and high efficiency. However, the target template of the model in both the PrDiMP and SiamBAN is not updated online, and feature vectors of the template image and the search image are independent of each other in the IoU-Net and Siamese frameworks. In this research, we proposed a template-guidance attention network in both the IoU-Net (denoted as TGAN-I) and Siamese (denoted as TGAN-S) frameworks for visual tracking. TGAN-I and TGAN-S can comprehensively utilize the feature information of the template image and search image, and provide an implicit way to update the template. By utilizing a simple template update strategy, the TGAN-I and TGAN-S trackers can be more robust under certain challenging conditions such as occlusion and deformation. Besides, we introduce a channel and spatial attention module in feature maps of the template image and search image for adaptive feature refinement. Deformable convolutional networks are further used to enhance the model generalization capability in various transformations aspect ratios and scales of tracking targets. To verify the effectiveness of the proposed method, we evaluate the TGAN-I and TGAN-S trackers on six benchmarks and achieve state-of-the-art results. In particular, the TGAN-I method outperforms the strong baseline, PrDiMP, by 0.323 → 0.355 and 0.471 → 0.501 of EAO score on VOT2019 and VOT2016, respectively.

Support vector regression for high-resolution beach surface moisture estimation from terrestrial LiDAR intensity data

Beach Surface Moisture (BSM) is a key attribute in the coastal investigations of land-atmospheric water and energy fluxes, groundwater resource budgets and coastal beach/dune development. In this study, an attempt has been made for the first time to estimate BSM from terrestrial LiDAR intensity data based on the Support Vector Regression (SVR). A long-range static terrestrial LiDAR (Riegl VZ-2000) was adopted to collect point cloud data of high spatiotemporal resolution on the Ostend-Mariakerke beach, Belgium. Based on the field moisture samples, SVR models were developed to retrieve BSM, using the backscattered intensity, scanning ranges and incidence angles as input features. The impacts of the training samples’ size and density on the predictive accuracy and generalization capability of the SVR models were fully investigated based on simulated BSM-intensity samples. Additionally, we compared the performance of the SVR models for BSM estimation with the traditional Stepwise Regression (SR) method and the Artificial Neural Network (ANN). Results show that SVR could accurately retrieve the BSM from the backscattered intensity with high reproducibility (average test RMSE of 0.71% ± 0.02% and R2 of 0.98% ± 0.002%). The Radial Basis Function (RBF) was the most suitable kernel for SVR model development in this study. The impacts of scanning geometry on the intensity could also be accurately corrected in the process of estimating BSM by the SVR models. However, compared to the SR method, the predictive accuracy and generalization performance of SVR models were significantly dependent on the training samples’ coverage, size and distribution, suggesting the need for the training samples of uniform distribution and representativeness. The minimum size of training samples required for SVR model development was 54. Under this condition, SVR performed similarly to ANN with a test RMSE of 1.06%, but SVR still performed acceptably (with an RMSE of 1.83%) even using extremely few training samples (only 16 field samples of uniform distribution), far better than the ANN (with an RMSE of 4.02%).

AUTOMATED BUILDING SEGMENTATION AND DAMAGE ASSESSMENT FROM SATELLITE IMAGES FOR DISASTER RELIEF

Abstract. After a natural disaster or humanitarian crisis, rescue forces and relief organisations are dependent on fast, area-wide and accurate information on the damage caused to infrastructure and the situation on the ground. This study focuses on the assessment of building damage levels on optical satellite imagery with a two-step ensemble model performing building segmentation and damage classification trained on a public dataset. We provide an extensive generalization study on pre- and post-disaster data from the passage of the cyclone Idai over Beira, Mozambique, in 2019 and the explosion in Beirut, Lebanon, in 2020. Critical challenges are addressed, including the detection of clustered buildings with uncommon visual appearances, the classification of damage levels by both humans and deep learning models, and the impact of varying imagery acquisition conditions. We show promising building damage assessment results and highlight the strong performance impact of data pre-processing on the generalization capability of deep convolutional models.

Efficient computational techniques for predicting the California bearing ratio of soil in soaked conditions

California bearing ratio (CBR) is one of the important parameters that is used to express the strength of the pavement subgrade of railways, roadways, and airport runways. CBR is usually determined in the laboratory in soaked conditions, which is an exhaustive and time-consuming process. Therefore, to sidestep the operation of conducting actual laboratory tests, this study presents the development of four efficient soft computing techniques, namely multivariate adaptive regression splines with piecewise linear models (MARS-L), multivariate adaptive regression splines with piecewise cubic models (MARS-C), Gaussian process regression, and genetic programming. For this purpose, a wide range of experimental results of soaked CBR was collected from an ongoing railway project of Indian Railways. Three explicit expressions are proposed to estimate the CBR of soils in soaked conditions. Separate laboratory experiments were performed to evaluate the generalization capabilities of the developed models. Furthermore, simulated datasets were used to validate the feasibility of the best-performing model. Experimental results reveal that the proposed MARS-L model attained the most accurate prediction (R2 = 0.9686 and RMSE = 0.0359 against separate laboratory experiments) in predicting the soaked CBR at all stages. Based on the accuracies attained, the proposed MARS-L model is very potential to be an alternate solution to estimate the CBR value in different phases of civil engineering projects.

Robust Deep Identification using ECG and Multimodal Biometrics for Industrial Internet of Things

The use of electrocardiogram (ECG) data for personal identification in Industrial Internet of Things can achieve near-perfect accuracy in an ideal condition. However, real-life ECG data are often exposed to various types of noises and interferences. A reliable and enhanced identification method could be achieved by employing additional features from other biometric sources. This work, thus, proposes a novel robust and reliable identification technique grounded on multimodal biometrics, which utilizes deep learning to combine fingerprint, ECG and facial image data, particularly useful for identification and gender classification purposes. The multimodal approach allows the model to deal with a range of input domains removing the requirement of independent training on each modality, and inter-domain correlation can improve the model generalization capability on these tasks. In multitask learning, losses from one task help to regularize others, thus, leading to better overall performances. The proposed approach merges the embedding of multimodality by using feature-level and score level fusions. To the best of our understanding, the key concepts presented herein is a pioneering work combining multimodality, multitasking and different fusion methods. The proposed model achieves a better generalization on the benchmark dataset used while the feature-level fusion outperforms other fusion methods. The proposed model is validated on noisy and incomplete data with missing modalities and the analyses on the experimental results are provided.

Comparative Study of Deep Generative Models on Chemical Space Coverage.

In recent years, deep molecular generative models have emerged as promising methods for de novo molecular design. Thanks to the rapid advance of deep learning techniques, deep learning architectures such as recurrent neural networks, variational autoencoders, and adversarial networks have been successfully employed for constructing generative models. Recently, quite a few metrics have been proposed to evaluate these deep generative models. However, many of these metrics cannot evaluate the chemical space coverage of sampled molecules. This work presents a novel and complementary metric for evaluating deep molecular generative models. The metric is based on the chemical space coverage of a reference dataset-GDB-13. The performance of seven different molecular generative models was compared by calculating what fraction of the structures, ring systems, and functional groups could be reproduced from the largely unseen reference set when using only a small fraction of GDB-13 for training. The results show that the performance of the generative models studied varies significantly using the benchmark metrics introduced herein, such that the generalization capabilities of the generative models can be clearly differentiated. In addition, the coverages of GDB-13 ring systems and functional groups were compared between the models. Our study provides a useful new metric that can be used for evaluating and comparing generative models.

Multi-Weight Domain Adversarial Network for Partial-Set Transfer Diagnosis

To realize fault identification of unlabeled data and improve model generalization capability, domain adaptation technology has been increasingly applied to intelligent fault diagnosis of machinery. Nevertheless, traditional domain adaptation diagnosis models generally restrict different domains to have the same label space, which does not always hold in complex industrial scenarios. Consequently, a more practical scenario, i.e., partial-set transfer diagnosis, is explored in this article, where the target label space is a subspace of the source domain. A multiweight domain adversarial network (MWDAN) is proposed to solve this issue, in which the weighting mechanism of class-level and instance-level is jointly designed to distinguish the label space and quantify the transferability of data samples. Based on the proposed strategy, the positive transfer between shared classes is promoted while the negative effect caused by outlier classes is circumvented. As a result, MWDAN can learn discriminative representations for accurate fault diagnosis in the target domain. Extensive experiments constructed on two mechanical systems demonstrate the outstanding performance of MWDAN.

Generalization performance of a deep learning based engine-out emissions model

In our previous work [1], an ensemble of Convolutional Neural Network (CNN) models was used to predict engine-out emissions of Carbon Monoxide (CO), Unburned Hydrocarbons (HC) and Smoke directly from in-cylinder contours of multiple scalar fields - Equivalence Ratio, Temperature, Velocity and Turbulent Kinetic Energy (TKE) at Exhaust Valve Opening (EVO) generated by physics based Computational Fluid Dynamics (CFD) simulations and experimentally measured emissions data. A dataset of 600 CFD generated images, one for each engine operating condition, from a 1.6L light-duty diesel engine was used for training, validation and testing of the CNN ensemble model. In this short communication, several modifications were made to the image data generation and training procedure. The ensemble of CNN models was re-trained per the new procedure on the modified image dataset from the 1.6L engine. The generalization capability of this re-trained deep learning model was investigated by evaluating its prediction performance on a test dataset from a 3.0L diesel engine that was not used previously for training, validation or testing. The CNN ensemble model re-trained on the 1.6L engine dataset captured the qualitative trends for all three emissions species on the 3.0L engine test dataset, however, quantitatively its prediction performance was not as good. Future work with additional engines and input features is planned to improve the generalization performance of the CNN based emissions prediction model.

Inverse design of unparametrized nanostructures by generating images from spectra.

Recently, there has been an increasing number of studies applying machine learning techniques for the design of nanostructures. Most of these studies train a deep neural network (DNN) to approximate the highly nonlinear function of the underlying physical mapping between spectra and nanostructures. At the end of training, the DNN allows an on-demand design of nanostructures, i.e., the model can infer nanostructure geometries for desired spectra. While these approaches have presented a new paradigm, they are limited in the complexity of the structures proposed, often bound to parametric geometries. Here we introduce spectra2pix, which is a DNN trained to generate 2D images of the target nanostructures. By predicting an image, our model architecture is not limited to a closed set of nanostructure shapes, and can be trained for the design of a much wider space of geometries. We show, for the first time, to the best of our knowledge, a successful generalization ability, by designing completely unseen shapes of geometries. We attribute the successful generalization to the ability of a pixel-wise architecture to learn local properties of the meta-material, therefore mimicking faithfully the underlying physical process. Importantly, beyond synthetical data, we show our model generalization capability on real experimental data.

Evolutionary artificial intelligence approach for performance prediction of bio-composites

Giving the high amount of carbon and energy emission from the use of traditional building materials, the use of bio-composites made from industrial crops especially hemp has caught attention from researchers in recent years. These bio-composites not only enhance the thermal performance of buildings but also promote sustainable development due to their eco-friendly nature. Due to their highly heterogeneous nature, however, most of the existing studies on the bio-composites have only focused on experimental investigations, while mathematical modeling of physical, thermal and mechanical properties of bio-composite remains a challenge for the researchers. In this paper, an artificial intelligence (AI) based gene expression programming (GEP) technique is used to develop the mathematical models for predicting the dry density, compressive strength and thermal conductivity of hemp-based bio-composites. A large amount of database was established based on past studies and the most influential parameters were identified by several trial analyses. The proposed mathematical models showed a high correlation with the experimental results. All the models passed the statistical and performance index checks showing strong predictability, generalization capability and high accuracy of GEP-AI models. Comparison of results with the regression analysis techniques further proved the superiority of GEP-AI models over the traditional methods.