Generalization Capacity Research Articles

Appearance-Pose Joint Coordinates Information Collaboration Model for clothes-changing person re-identification

Clothes-changing person re-identification (CC-ReID) is of great importance in the field of intelligent surveillance as it can be used to help police officers find criminals who want to evade pursuit by changing clothes. Currently, the best-performing algorithm overly emphasizes discriminative information in appearance, which limits the generalization capability of CC-ReID. In this paper, we innovatively propose an Appearance-Pose Joint Coordinate Information Collaboration Model (AP-ICM), which further enhances the generalization capability of the CC-ReID method. FIRST, the pose joint coordinates, a novel auxiliary information that encodes the body shape, is introduced. Compared to the existing auxiliary information ( i.e. contour sketches, human masks, heatmaps) used to encode body shape, it avoids the interference of changes of clothing shape and the interference from areas unrelated to the joints, thus the model can learn discriminative information from the body shape more efficiently. Experiments demonstrate that the pose joint coordinates is more effective than these existing auxiliary information in assisting the model to learn discriminative information from the body shape. THEN, a novel Second-order Structural Information is designed to help the model learn the discriminative information from the pose joint coordinates more easily, thus improving the generalization capability. Experiment demonstrate that the Second-order Structural Information is effective in enhancing generalization capacity. NEXT, a novel Semantic Representation endows the model with the ability to understand semantic differences between joints, so that the model can learn features more accurately during training, which in turn makes it easier to acquire strong generalization capability. Experiment demonstrate that the Semantic Representation is effective in enhancing generalization capacity. FINALLY, by adopting a Information Collaboration Module, we successfully integrate discriminative information extracted from the pose joint coordinates into the best-performing algorithm. Experiments demonstrate that our AP-ICM achieves Rank-1 of 68.50% and 82.90% on two generic CC-ReID datasets, PRCC and VC-Clothes, respectively, outperforming the best-performing algorithm by 2.30% and 2.90%, respectively.

New chaos-integrated improved grey wolf optimization based models for automatic detection of depression in online social media and networks

Depression is a psychological effect of the modern lifestyle on people’s thoughts. It is a serious individual and social health problem due to the risk of suicide and loss of workforce, high chronicity, recurrence rates, and prevalence. Therefore, identification, prevention, treatment of depression, and determination of relapse risk factors are of great importance. Depression has traditionally been diagnosed using standardized scales that require clinical diagnoses or patients’ subjective responses. However, these classical techniques have some limitations such as cost, uncomfortability, subjectivity, and ineffectiveness. Social media data can be simply and efficiently used for depression detection because it allows instantaneous emotional expression and quick access to various information. Some machine learning-based methods are used for detecting the depression in online social media and networks. Nevertheless, these algorithms suffer from several drawbacks, including data sparsity, dimension explosion, restricted capacity for generalization, and low performance on imbalanced data sets. Furthermore, many machine learning methods work as black-box models, and the constructed depression detection models are not interpretable and explainable. Intelligent metaheuristic optimization algorithms are widely used for different types of complex real-world problems due to their simplicity and high performance. It is aimed to remove the limitations of studies on this problem by increasing the success rate and automatically selecting the relevant features and integrating the explainability. In this study, new chaos-integrated multi-objective optimization algorithms are proposed to increase efficiency. New improved Grey Wolf Optimization algorithms have been proposed by integrating Circle, Logistic, and Iterative chaotic maps into the improved Grey Wolf Optimization algorithm. It is aimed to increase the success rate by proposing a multi-objective fitness function that can optimize the accuracy and the number of features simultaneously. The proposed algorithms are compared with different types of popular supervised machine learning algorithms and current metaheuristic algorithms that are widely and successfully used in depression detection problems. Experimental results show that the proposed models outperform machine learning methods, as evidenced by examining results with accuracy, F-measure, MCC, sensitivity, and precision measures. An accuracy value of 100% was obtained from proposed algorithms. In addition, when the confusion matrices are examined, it is seen that they exhibit a successful distribution. Although it is a new research and application area for optimization theory, promising results have been obtained from the proposed models.

VC dimensions of group convolutional neural networks

We study the generalization capacity of group convolutional neural networks. We identify precise estimates for the VC dimensions of simple sets of group convolutional neural networks. In particular, we find that for infinite groups and appropriately chosen convolutional kernels, already two-parameter families of convolutional neural networks have an infinite VC dimension, despite being invariant to the action of an infinite group.

A Generative Adversarial Network to Synthesize 3D Magnetohydrodynamic Distortions for Electrocardiogram Analyses Applied to Cardiac Magnetic Resonance Imaging.

Recently, deep learning (DL) models have been increasingly adopted for automatic analyses of medical data, including electrocardiograms (ECGs). Large, available ECG datasets, generally of high quality, often lack specific distortions, which could be helpful for enhancing DL-based algorithms. Synthetic ECG datasets could overcome this limitation. A generative adversarial network (GAN) was used to synthesize realistic 3D magnetohydrodynamic (MHD) distortion templates, as observed during magnetic resonance imaging (MRI), and then added to available ECG recordings to produce an augmented dataset. Similarity metrics, as well as the accuracy of a DL-based R-peak detector trained with and without data augmentation, were used to evaluate the effectiveness of the synthesized data. Three-dimensional MHD distortions produced by the proposed GAN were similar to the measured ones used as input. The precision of a DL-based R-peak detector, tested on actual unseen data, was significantly enhanced by data augmentation; its recall was higher when trained with augmented data. Using synthesized MHD-distorted ECGs significantly improves the accuracy of a DL-based R-peak detector, with a good generalization capacity. This provides a simple and effective alternative to collecting new patient data. DL-based algorithms for ECG analyses can suffer from bias or gaps in training datasets. Using a GAN to synthesize new data, as well as metrics to evaluate its performance, can overcome the scarcity issue of data availability.

Facial expression recognition based on ResNet and transfer learning

With its potential to revolutionize a wide range of applications, including lie detection, social robotics, and driver fatigue detection, facial expression recognition is a field that is rapidly expanding. However, traditional machine learning methods have struggled with facial expression recognition due to limitations such as manual feature selection and limited representation capabilities. Additionally, these methods require large amounts of annotated data, which can be time-consuming and expensive to obtain. In order to overcome these difficulties, this paper suggests a novel method that builds recognition models using a multi-layer perceptron (MLP) and ResNet. This hybrid model offers improved performance over conventional CNN models, achieving an impressive accuracy rate of 85.71% on the FER_2013 dataset. Additionally, migration learning is used to increase the model's precision and avoid over-fitting. The FER_2013 dataset is used to train and test the model. The results of the trials show that the suggested model can recognize facial expressions while minimizing the overfitting problem typically associated with deep learning. The model will eventually include a self-attentive mechanism in the study in an effort to improve model resolution. By using it with color images, the team also hopes to increase the model's capacity for generalization.

Effect of the deep residual networks at different depths on expression recognition

Facial expression recognition plays a critical role in numerous applications like emotion analysis, human-computer interaction, and surveillance systems. Given the importance of this task, this study aims to investigate the effectiveness of different depths of Residual Networks (ResNet). The primary objective is to scrutinize and compare these ResNet models in terms of their training and validation losses and performance metrics like accuracy, recall, and F1 scores. In this research, a thorough comparative analysis is conducted by setting up exhaustive experiments using these models. The experiment is carried out on a popular facial expression dataset. Despite the depth differences, ResNet101 emerged as the model demonstrating superior performance. It struck the most effective balance between model complexity and generalization capacity, leading to the lowest validation loss and better performance. Experimental results show that a more complex model does not necessarily yield optimal results. The optimal balance between model complexity and generalisation needs to be investigated. These findings can provide essential guidance in the design of deep learning models for facial expression recognition and other similar tasks.

Prediction of landslide step-like displacement using factor preprocessing-based hybrid optimized SVR model in the Three Gorges Reservoir, China

Artificial intelligence (AI) is gaining popularity in landslide displacement modeling due to its strong generalization capacity and ability to accurately represent complex, nonlinear laws. However, few studies have examined the validity of factor selection and the impact of data noise on the model. To address these issues, we employ various analyses, including Pearson correlation coefficient and grey correlation degree analyses to diagnose redundant and highly correlated factors, respectively. Singular spectrum analysis (SSA) is employed for denoising the data, leveraging the benefits of the hankel matrix and minimal parameters. This technique facilitates the decomposition of displacement into trend and periodic terms. The trend term displacement is fitted with a polynomial, and the periodic term displacement is predicted using a support vector regression (SVR) model. The hybrid-optimized SVR by particle swarm optimization-gravitational search algorithm (PSO-GSA) is used to improve the model's performance and prediction ability. We apply the proposed method to the Bazimen and Baishuihe landslides in Zigui County for model validation. The results show that denoising the input factor data reduces the root mean square error (RMSE) and mean absolute error (MAE) and increases the correlation coefficient (R). Furthermore, the PSO-GSA algorithm exhibits excellent generalization ability and fast optimization speed in optimizing parameters. The correlation coefficient increase to 0.99 when using the hybrid-optimized SVR model for displacement prediction.

An ensemble approach for enhancing generalization and extendibility of deep learning facilitated by transfer learning: principle and application in curing monitoring

Machine learning (ML) and deep learning (DL) have exhibited significant advantages compared to conventional data analysis methods. However, the limitations of poor generalization and extendibility impede the broader application of these methods beyond specific learning tasks. To address this challenge, this study proposes a transfer learning-based ensemble approach called SMART. This approach incorporates synthetic minority oversampling technique, average reinforced interpolation, series data imaging, and fine-tuning. To validate the effectiveness of SMART, we conduct experiments on curing monitoring of polymeric composites and construct a hybrid dataset with highly heterogeneous features. We compare the performance of SMART with exemplary ML algorithms using conventional evaluation indicators, including Accuracy, Precision, Recall, and F1-score. The experimental results demonstrate that the SMART approach exhibits superior generalization capacity and extendibility, achieving indicator scores above 0.9900 in new scenarios. These findings suggest that the proposed SMART approach has the potential to break through the limitations of conventional ML and DL models, enabling wider applications in the industrial sectors.

Deep learning for accelerating Radon inversion in single-cells tomographic phase imaging flow cytometry

Tomographic phase microscopy (TPM) in flow cytometry is an emerging imaging technology completely stain-free and able to provide unique 3D biophysical data. However, the required throughput for TPM would correspond to a very huge amount of volumetric data to be processed. To date, the most accurate Radon inversion algorithms for tomographic reconstructions are based on iterative methods aided by regularizations, thus providing high performance but at the cost of a demanding computation time. To balance the trade-off between reconstruction performance and reconstruction speed, learning-based approaches have been successfully introduced, thus demonstrating high accuracy in solving the Radon inversion problem. However, the complexity of the proposed models remains a bottleneck due to the high computational resources still required for the training phase. Here we show that, employing a multi-scale fully convolutional Context Aggregation Network (CAN) model, a significant speeding-up of the Radon inversion computation can be achieved. Compared to other conventional encoder-decoder networks such as U-Net, the proposed method has proven to be accurate, faster and particularly suitable for on-board processing. Moreover, we show the generalization capacity of CAN by training the network with simulated tomographic data and testing the learned model on experimental tomographic data.

DenSplitnet: Classifier-invariant neural network method to detect COVID-19 in chest CT data

Objective:COVID-19 has made an unprecedented impact on humanity. The Healthcare sector, in an effort to curb COVID-19, could leverage Artificial Intelligence (AI) to its aid, especially in diagnosing it through the classification of Chest-CT scans. However, data scarcity plagues the medical domain. Therefore, any AI solution must be capable of learning from limited data. Also the resulting AI could relieve radiologists from exhaustion and aid medical practitioners as a valuable diagnostic tool. Methods:Our proposed model DenSplitnet uses Dense blocks to learn image features, Self-Supervised Learning for pre-training to learn the context, and a novel two-way split branch at the final classification layer for classifier-invariant generalization ability. Results:DenSplitnet achieves state-of-the-art performance on four benchmark chest-CT scan datasets for COVID-19. The model performs well on the Sars-cov-2 ct-scan dataset. The Test accuracy is 91.92%, F1-Score is 91.30%, Precision is 96.55%, and Recall is 87.04%. The model achieves test accuracy of 86.17%, Precision of 82.94%, Recall of 83.72%, and F1-Score of 83.23% on the Sars-cov-2 ct-scan multiclass dataset. The model obtains a test accuracy of 86.21%, an F1-Score of 83.91%, and an AUC of 0.95 in the Covid-ct-dataset. The model obtains a test accuracy that is 73.11% and an F1-Score of 70.97% in the TransferLearn-ct dataset. Conclusion:The findings support the practical usability of the DenSplitnet as a technological AI assistance to radiologists and other medical professionals, alongside Grad-CAM plots for explainable AI and the theoretical examination of the classifier-invariant generalization capacity of the network. The healthcare sector can benefit from this technology in a number of ways.

Deep residual convolutional neural Network: An efficient technique for intrusion detection system

The fast growth of computer networks over the past few years has made network security in smart cities a significant issue. Network intrusion detection is crucial to maintaining the integrity, confidentiality, and resource accessibility of the various network security rules. Conventional intrusion detection systems frequently use mining association rules to identify intrusion behaviors. They run into issues such as a high false alarm rate (FAR), limited generalization capacity, and slow timeliness because they cannot adequately extract distinctive information about user activities. The primary goal of the current research is to classify attacks using efficient approaches to identify genuine packets. If the number of characteristics in a dataset decreases, the complexity of DL approaches is significantly decreased. In this research work, the Deep Residual Convolutional neural network (DCRNN) is proposed to enhance network security through intrusion detection, which is optimized by the Improved Gazelle Optimization Algorithm (IGOA). Feature selection has eliminated irrelevant features from network data used in obstacle classification processes. Essential features are chosen using the Novel Binary Grasshopper Optimization Algorithm (NBGOA). Experimentation is carried out using the UNSW-NB-15, Cicddos2019 dataset, and CIC-IDS-2017 dataset. According to the experimental findings, the proposed system outperforms existing models regarding classification accuracy and processing time. The results demonstrate that the presented approach efficiently and precisely identifies various assaults.

Bayesian Convolutional Neural Networks in Medical Imaging Classification: A Promising Solution for Deep Learning Limits in Data Scarcity Scenarios

Deep neural networks (DNNs) have already impacted the field of medicine in data analysis, classification, and image processing. Unfortunately, their performance is drastically reduced when datasets are scarce in nature (e.g., rare diseases or early-research data). In such scenarios, DNNs display poor capacity for generalization and often lead to highly biased estimates and silent failures. Moreover, deterministic systems cannot provide epistemic uncertainty, a key component to asserting the model’s reliability. In this work, we developed a probabilistic system for classification as a framework for addressing the aforementioned criticalities. Specifically, we implemented a Bayesian convolutional neural network (BCNN) for the classification of cardiac amyloidosis (CA) subtypes. We prepared four different CNNs: base-deterministic, dropout-deterministic, dropout-Bayesian, and Bayesian. We then trained them on a dataset of 1107 PET images from 47 CA and control patients (data scarcity scenario). The Bayesian model achieved performances (78.28 (1.99) % test accuracy) comparable to the base-deterministic, dropout-deterministic, and dropout-Bayesian ones, while showing strongly increased “Out of Distribution” input detection (validation-test accuracy mismatch reduction). Additionally, both the dropout-Bayesian and the Bayesian models enriched the classification through confidence estimates, while reducing the criticalities of the dropout-deterministic and base-deterministic approaches. This in turn increased the model’s reliability, also providing much needed insights into the network’s estimates. The obtained results suggest that a Bayesian CNN can be a promising solution for addressing the challenges posed by data scarcity in medical imaging classification tasks.

Debiased Contrastive Curriculum Learning for Progressive Generalizable Person Re-Identification

Domain generalization (DG) in person re-identification (ReID) is an extremely challenging but essential task, which aims to learn a generalizable model over multiple labeled source domains that can perform well on unseen target domains. Most existing DG strategies in ReID directly aggregate multiple source data together for training, incurring a large inter-domain bias and unstable model optimization that lead the model apt to overfitting domain bias and the model training more time-consuming respectively, thus hampering the generalization and convergence speed of the model. To tackle the aforementioned issues, inspired by Curriculum Learning that mimics the process of human lifelong DG learning (from easy to hard), we put forward a novel Debiased Contrastive Curriculum Learning (DCCL) strategy for DG ReID, which is designed to incrementally enhance generalization in an easy-to-hard training way that can continuously accumulate learning experience to make learning in unknown domains easier and effectively eliminate the domain bias to help the model learn rich domain-invariant discriminative features, thereby strengthening generalization and accelerating convergence for the model. In addition, to simultaneously learn class-level and instance-level discriminative representations, we raise a non-parametric hybrid contrastive loss to equip the DCCL model. We also particularly design an inter-domain mix module to variegate the features of the newly added source domain at each stage of DCCL, further establishing the advantages of DCCL. Extensive experimental results on four public ReID benchmarks fully demonstrate that our DCCL can effectively strengthen the generalization capacity of the model to unseen domains and outperform the state-of-the-art methods.

RCNet: Related Context-Driven Network with Hierarchical Attention for Salient Object Detection

Recent progress in salient object detection (SOD) mainly depends on dilated convolution with different receptive fields to capture contextual information for multi-scale learning. Intuitively, contextual information in different scales is conducive to understanding the image content, and thus can help us identify and locate salient objects in real-world scenes. However, the sparsity inside the dilated convolution kernel may cause the problem of local information loss, limiting the predictive accuracy of the model. In addition, the inequality of feature channels should also be considered, and they often feature different deviations for salient objects or background noises. Although some channel attention mechanisms have been proposed in SOD, their ability to capture global information is limited, and the problem of high complexity is still a great challenge. To alleviate the abovementioned problems, we propose a Related Context-Driven Network (RCNet) with Hierarchical Attention for Salient Object Detection, consisting of a cascaded multi-scale context exploration (CMCE) module and a hierarchical feature aggregation (HFA) module. The CMCE module is to capture multi-scale contextual information through using multi-receptive-field dilated convolutions in a diamond hierarchical structure, where a feature reconstruction operation is deployed to improve the correlation of features, effectively avoiding the gridding problems and local information loss. Meanwhile, the HFA module adaptively interacts with the complementary information of the multi-level features to further capture the important information from within the feature channel by a multi-source hybrid channel attention (MHCA) mechanism to generate powerful and robust feature representations. Extensive experiments on six benchmark datasets demonstrate that the proposed RCNet method consistently outperforms 20 existing the state-of-the-art SOD methods in terms of accuracy, generalization capacity and robustness.

System for quantitative diagnosis of COVID-19 associated Pneumonia based on Superpixels with deep learning and chest CT

COVID-19 is a disease caused by the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) that can lead to complications such as acute respiratory distress syndrome, acute heart injury and secondary infections in a relatively high proportion of patients and, consequently, lead to significant mortality. The definitive diagnosis of COVID-19 is performed by real-time Polymerase Chain Reaction (RT-PCR). However, as RT-PCR results usually take longer to be completed than computed tomography (CT), the latter has had an important role in the detection of patients infected with COVID-19. A rough estimate of the extent of lung involvement by the disease is also important and considered an additional criterion for deciding on discharge or hospitalization. Recent research has adopted deep neural networks and other machine learning approaches to detect the presence of lung infection caused by COVID-19. However, the extent of lung involvement (volume) caused by the disease has been only superficially investigated. In this work, we created an end-to-end computer vision system to automatically quantify the Percentage of Infection (POI) in chest CT images of COVID-19 cases confirmed by laboratorial analysis. We evaluated the performance of three well-known deep neural networks, trained using three different strategies: (1) no transfer learning with randomly initialized weights; (2) transfer learning using with ImageNet weights without fine-tuning; and (3) transfer learning with weight fine-tuning. Data augmentation and dropout were used during training to reduce overfitting and increase the generalization capacity of the models. Our approach consists of segmenting a chest CT with the SLIC Superpixels method and classifying each segment (superpixel) into a specific class (COVID or non-COVID). We used the weights of the best deep neural network to classify the superpixels and quantify the COVID-19 infection by calculating the POI on chest CT. The results indicate that deep learning models can be successfully used to support radiologists in the quantitative diagnosis of lung infection caused by COVID-19, reaching an accuracy of up to 98.4% with Inception-Resnet-v2 architecture.

Learning to Class-Adaptively Manipulate Embeddings for Few-Shot Learning

In few-shot learning (FSL), meta-learning approach (MLA) mainly focuses on learning transferable knowledge from plenty of auxiliary FSL tasks to facilitate fast generalization to a new task. For a given FSL task, due to the inter-class distribution discrepancy, each class necessitates a specific embedding (i.e., a mapping function) to map samples into an ideal semantic space where samples from this class can be well separately from other classes. Moreover, these embeddings may vary with different tasks. Hence, one crucial knowledge for MLA is how to separately construct optimal embeddings for each class based on a few training samples given in a FSL task. However, most existing MLAs rarely consider this and thus show limited generalization capacity. To mitigate this problem, instead of directly construct class-adaptive embeddings, we present a new MLA that aims at learning to class-adaptively manipulate the features of samples for accurate classification in a new FSL task. In a specific, for a new FSL task, the proposed MLA first learns to generate some class-specific weights based on training samples via exploiting the inter-class distribution discrepancy between this class and the others. Then, the generated weights are utilized to compute the Hadamard product of features produced by a task-agnostic embedding module. By doing this, the proposed MLA can dynamically enhance or depress some specific semantic dimensions of sample features depending on the distribution of each class for accurate classification, and thus equals to constructing class-adaptive embeddings for each class but in a simpler way which can appropriately avoid over-fitting and is scalable to cases with extensive classes. To show its efficacy, we test the proposed MLA on four benchmark FSL datasets under various settings and report superior performance over existing state-of-the-arts.

High-Dimensional Data Global Sensitivity Analysis Based on Deep Soft Sensor Model.

This article investigates the sensitivity analysis (SA) of high-dimensional data to identify the effects of process variables on output quantity of interest (QoI) in industrial soft sensor modeling. The computational cost of analyzing the SA of high-dimensional data is high, and models available for SA techniques usually have limited generalization capacity. Therefore, we propose a novel high-dimensional data global SA (GSA) approach based on a deep soft sensor model to address these issues. We first develop an approximately incremental grouping (AIG) algorithm and a region-based cooperative co-evolution (RBCC) algorithm to decompose the high-dimensional data into independent regions for the GSA. Subsequently, a multihead deep soft sensor model with generalization performance is designed to determine the GSA indices of each decomposed region. Specifically, the region of interest (RoI) align algorithm provides the multihead with precisely located decomposed region features. Finally, based on the uncertainty analysis of each model head, we present a joint loss function with the Monte Carlo dropout (MC-dropout) algorithm to measure the GSA indices of each decomposed region on QoIs. Experimental evaluation results on a benchmark dataset and a real-world one demonstrate the effectiveness of the proposed approach in addressing the GSA of high-dimensional data in industrial processes.

Modeling of water surface profile in non-prismatic compound channels

Abstract Estimating the water surface elevation of river systems is one of the most complicated tasks in formulating hydraulic models for flood control and floodplain management. Consequently, utilizing simulation models to calibrate and validate the experimental data is crucial. HEC-RAS is used to calibrate and verify the water surface profiles for various converging compound channels in this investigation. Based on experimental data for converging channels (θ = 5°, 9°, and 12.38°), two distinct flow regimes were evaluated for validation. The predicted water surface profiles for two relative depths (β = 0.25 and 0.30) follow the same variational pattern as the experimental findings and are slightly lower than the observed values. The MAPE for the simulated and experimental results is less than 3%, indicating the predicted HEC-RAS value performance and accuracy. Therefore, our findings imply that in the case of non-prismatic rivers, the proposed HEC-RAS models are reliable for predicting water surface profiles with a high generalization capacity and do not exhibit overtraining. However, the results demonstrated that numerous variables impacting the water surface profile should be carefully considered since this would increase the disparities between HEC-RAS and experimental data.

Industrial UAV-Based Unsupervised Domain Adaptive Crack Recognitions: From Database Towards Real-Site Infrastructural Inspections

The defect diagnosis of modern infrastructures is crucial to public safety. In this work, we propose an unsupervised domain adaptive crack recognition framework. To fulfill the unsupervised domain adaptation (UDA) task of cracks recognition in infrastructural inspections, we propose a robust unsupervised domain adaptive learning strategy termed <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Crack-DA</i> to increase the generalization capacity of the model in unseen test circumstances. More specifically, we first propose leveraging the self-supervised depth information to help the learning of semantics. And then using the edge information to suppress non-edge background objects and noises. We propose the data augmentation-based consistency to enhance the performance. More importantly, we proposed to use the disparity in depth to evaluate the domain gap in semantics and explicitly consider the domain gap in network optimization. A database composed of a large number of crack images with detailed pixel-level labels for network training is established. Extensive experiments on UAV-captured highway cracks and real-site UAV inspections of building cracks demonstrate the robustness and effectiveness of our proposed domain adaptive crack recognition approach.

Heating energy demand estimation of the EU building stock: Combining building physics and artificial neural networks

The aim of this study is to present a novel data-driven approach developed for space heating energy demand calculation of the whole EU building stock. To develop a computationally efficient bottom-up model that takes into account building physics parameters and details of the building stock make-up, an artificial neural network (ANN) is trained on a dataset of precise building-physics models. For this purpose, 2025 building energy simulations were performed in this study, ensuring representativeness for the entire EU, in terms of building stock properties (3 building geometries and 27 combinations of building envelope properties) and climatic conditions (25 weather files). The developed model is the first comprehensive hybrid (ANN and building physics) approach based on a condensed selection of input features, yet covering a wide range of climatic and building stock values, hence suitable for consideration of the spatial, temporal and multinational scale of the EU building stock. A unique building stock model was developed for this study at the EU regional level. The performance of the ANNs was validated by assessing the RMSE and R2 against the training (85%), testing (15%) and total dataset (100%, i.e. 2025 data points), to ensure adequate robustness, accuracy and generalisation capacity. As to the application of the ANN model on the building stock of 27 EU countries, a good accordance between ANN model results and measured consumption over the nine years from 2011 to 2019, was obtained, with an average overestimation by 9.3% (standard deviation (SD) = 5%), which confirmed the adequacy of the adopted methodology. Additional consideration of the occupant behaviour resulted in average overestimating the actual measurements by only 2.4% (SD = 4.7%). In addition to the EU and country level, the added value of the implemented bottom-up approach is its ability to assess the heating energy demand at regional level. Overall, the presented approach can be considered accurate and reliable to be used in future research on the long-term impact of building renovation with respect to energy savings and reductions of greenhouse gas emissions.