Generalization Capability Research Articles

Research and Application of Traffic Flow Prediction in Chengdu City Based on Neural Network Algorithms

As urban traffic systems increase in complexity, accurately predicting traffic flow has become a critical task in traffic management and planning. This paper employs neural network technologies, specifically Graph Convolutional Networks (GCN) and Long Short-Term Memory networks (LSTM), to develop a traffic flow prediction model for Chengdu that incorporates spatio-temporal characteristics. Through empirical analysis, the model demonstrates high accuracy and good generalization capability, effectively predicting and analyzing traffic flow at different times and regions. Additionally, this study explores the application of the prediction model in traffic law regulation and policy formulation, providing data support and scientific basis for advancing smarter traffic systems.

Efficient multi-fidelity reduced-order modeling for nonlinear flutter prediction

Flutter prediction is an important part of aircraft design. However, high-fidelity predictions for transonic flutter are difficult to make because of the associated computational costs. This paper proposes a multi-fidelity reduced-order modeling (MFROM) framework for flutter predictions to achieve high-fidelity simulations with limited computational costs. The high-fidelity data were obtained from a Navier–Stokes-equation-based solver, while the low-fidelity data were taken from an Euler-equation-based flow solver. By employing a multi-fidelity neural network trained with two types of data, this methodology enables nonlinear predictions for transonic results. To demonstrate the multi-fidelity process, a widely used pitching and plunging airfoil case is considered. Verification of the approach was performed by comparing with results from time-domain aeroelastic solvers. The results showed that the proposed multi-fidelity neural network modeling framework could realize online predictions of unsteady aerodynamic forces and flutter responses across multiple Mach numbers. Compared with the typical multi-fidelity method, the proposed neural network has a higher accuracy and a stronger generalization capability. Finally, the potential of the method to reduce the computational effort of high-fidelity aeroelastic analysis was demonstrated.

Scooby: Modeling multi-modal genomic profiles from DNA sequence at single-cell resolution.

Understanding how regulatory DNA elements shape gene expression across individual cells is a fundamental challenge in genomics. Joint RNA-seq and epigenomic profiling provides opportunities to build unifying models of gene regulation capturing sequence determinants across steps of gene expression. However, current models, developed primarily for bulk omics data, fail to capture the cellular heterogeneity and dynamic processes revealed by single-cell multi-modal technologies. Here, we introduce scooby, the first model to predict scRNA-seq coverage and scATAC-seq insertion profiles along the genome from sequence at single-cell resolution. For this, we leverage the pre-trained multi-omics profile predictor Borzoi as a foundation model, equip it with a cell-specific decoder, and fine-tune its sequence embeddings. Specifically, we condition the decoder on the cell position in a precomputed single-cell embedding resulting in strong generalization capability. Applied to a hematopoiesis dataset, scooby recapitulates cell-specific expression levels of held-out genes and cells, and identifies regulators and their putative target genes through in silico motif deletion. Moreover, accurate variant effect prediction with scooby allows for breaking down bulk eQTL effects into single-cell effects and delineating their impact on chromatin accessibility and gene expression. We anticipate scooby to aid unraveling the complexities of gene regulation at the resolution of individual cells.

Unified View Empirical Study for Large Pretrained Model on Cross-Domain Few-Shot Learning

The challenge of cross-domain few-shot learning (CD-FSL) stems from the substantial distribution disparities between target and source domain images, necessitating a model with robust generalization capabilities. In this work, we posit that large-scale pretrained models are pivotal in addressing the CD-FSL task owing to their exceptional representational and generalization prowess. To our knowledge, no existing research comprehensively investigates the utility of large-scale pretrained models in the CD-FSL context. Addressing this gap, our study presents an exhaustive empirical assessment of the Contrastive Language–Image Pre-Training model within the CD-FSL task. We undertake a comparison spanning six dimensions: base model, transfer module, classifier, loss, data augmentation, and training schedule. Furthermore, we establish a straightforward baseline model, E-base, based on our empirical analysis, underscoring the importance of our investigation. Experimental results substantiate the efficacy of our model, yielding a mean gain of 1.2% in 5-way 5-shot evaluations on the BSCD dataset.

RDAG U-Net: An Advanced AI Model for Efficient and Accurate CT Scan Analysis of SARS-CoV-2 Pneumonia Lesions.

Background/Objective: This study aims to utilize advanced artificial intelligence (AI) image recog-nition technologies to establish a robust system for identifying features in lung computed tomog-raphy (CT) scans, thereby detecting respiratory infections such as SARS-CoV-2 pneumonia. Spe-cifically, the research focuses on developing a new model called Residual-Dense-Attention Gates U-Net (RDAG U-Net) to improve accuracy and efficiency in identification. Methods: This study employed Attention U-Net, Attention Res U-Net, and the newly developed RDAG U-Net model. RDAG U-Net extends the U-Net architecture by incorporating ResBlock and DenseBlock modules in the encoder to retain training parameters and reduce computation time. The training dataset in-cludes 3,520 CT scans from an open database, augmented to 10,560 samples through data en-hancement techniques. The research also focused on optimizing convolutional architectures, image preprocessing, interpolation methods, data management, and extensive fine-tuning of training parameters and neural network modules. Result: The RDAG U-Net model achieved an outstanding accuracy of 93.29% in identifying pulmonary lesions, with a 45% reduction in computation time compared to other models. The study demonstrated that RDAG U-Net performed stably during training and exhibited good generalization capability by evaluating loss values, model-predicted lesion annotations, and validation-epoch curves. Furthermore, using ITK-Snap to convert 2D pre-dictions into 3D lung and lesion segmentation models, the results delineated lesion contours, en-hancing interpretability. Conclusion: The RDAG U-Net model showed significant improvements in accuracy and efficiency in the analysis of CT images for SARS-CoV-2 pneumonia, achieving a 93.29% recognition accuracy and reducing computation time by 45% compared to other models. These results indicate the potential of the RDAG U-Net model in clinical applications, as it can accelerate the detection of pulmonary lesions and effectively enhance diagnostic accuracy. Additionally, the 2D and 3D visualization results allow physicians to understand lesions' morphology and distribution better, strengthening decision support capabilities and providing valuable medical diagnosis and treatment planning tools.

Deep diffusion learning of mutual-reflective structured light patterns for multi-body three-dimensional imaging

Simultaneous structured light imaging of multiple objects has become more demanding and widely in many scenarios involving robot operations in intelligent manufacturing. However, it is challenged by pattern aliasing caused by mutual reflection between high-reflective objects. To this end, we propose to learn clear fringe patterns from aliased mutual-reflective observations by diffusion models for achieving high-fidelity multi-body reconstruction in line with typical phase-shift algorithms. Regarding mutual reflection imaging as a formation of adding significant noise, we build a supervised generative learning framework based on diffusion models and then train a self-attention-based deep network with a U-Net-like skip-connected encoder-decoder architecture. We demonstrate the generalization capability of the trained model in fringe pattern recovery and its performance in phase and three-dimensional (3D) shape reconstruction. Both experimental results show that the proposed method has the expected feasibility and accuracy, heralding a promising solution for addressing the current challenge in various multi-body mutual-reflective 3D reconstruction tasks.

Real-Time ConvNext-Based U-Net with Feature Infusion for Egg Microcrack Detection

Real-time automatic detection of microcracks in eggs is crucial for ensuring egg quality and safety, yet rapid detection of micron-scale cracks remains challenging. This study introduces a real-time ConvNext-Based U-Net model with Feature Infusion (CBU-FI Net) for egg microcrack detection. Leveraging edge features and spatial continuity of cracks, we incorporate an edge feature infusion module in the encoder and design a multi-scale feature aggregation strategy in the decoder to enhance the extraction of both local details and global semantic information. By introducing large convolution kernels and depth-wise separable convolution from ConvNext, the model significantly reduces network parameters compared to the original U-Net. Additionally, a composite loss function is devised to address class imbalance issues. Experimental results on a dataset comprising over 3400 graded egg microcrack image patches demonstrate that CBU-FI Net achieves a reduction in parameters to one-third the amount in the original U-Net, with an inference speed of 21 ms per image (1 million pixels). The model achieves a Crack-IoU of 65.51% for microcracks smaller than 20 μm and a Crack-IoU and MIoU of 60.76% and 80.22%, respectively, for even smaller cracks (less than 5 μm), achieving high-precision, real-time detection of egg microcracks. Furthermore, on the publicly benchmarked CrackSeg9k dataset, CBU-FI Net achieves an inference speed of 4 ms for 400 × 400 resolution images, with an MIoU of 81.38%, proving the proposed method’s robustness and generalization capability across various cracks and complex backgrounds.

MH-DCNet: An improved flow field prediction framework coupling neural network with physics solver

With the development of intelligent computing technology, deep learning methods have provided an efficient solution for rapid flow field prediction in computational fluid dynamics (CFD) problems. However, existing methods have limitations in handling interference among physical variables due to different data distributions, leading to a decline in prediction performance. In this paper, we propose MH-DCNet, an improved flow field prediction framework that couples a neural network with a physics solver. Specifically, to address the data distribution problem, we design a multi-head deep convolutional neural network that decouples the prediction of physical variables through multiple encoders and decoders. We also develop a hybrid loss function by introducing the mean structural similarity to better capture the complex spatial structures and distribution features of flow fields. We evaluate MH-DCNet with unseen geometries and various flow conditions. Experimental results show that MH-DCNet outperforms other advanced models in efficiency and generalization capability. It accelerates the prediction process by 2.35 times compared to the CFD method while meeting the convergence constraints.

To compete or strategically retreat? The global diffusion of reconnaissance strike

The reconnaissance strike complex is synonymous with modern military power, and prominent realist theories would have predicted rapid proliferation after its successful debut in the Gulf War. Instead, the complex has proliferated slowly. To explain this puzzle, we theorize that interstate security threats significantly impact proliferation, but not in the way traditionally presumed. Although the literature on weapons proliferation has largely assumed a monotonically increasing relationship should hold between the capabilities of a state’s adversaries and a state’s own capability, we draw from the economics literature and game theoretic insights from political science to argue that the relationship should resemble an inverted-U. When states have rivals with moderate reconnaissance strike capabilities, they have security incentives to compete with them. However, when states face highly advanced adversaries, it becomes more difficult to escape or match their competition, making symmetrical acquisition less appealing. While most prior research focuses on narrower aspects of the reconnaissance strike complex like missiles or smart bombs, we test our theory on a novel dataset tracking country-level acquisition of eight aspects of the complex from 1980 to 2017: Ballistic missiles; bombers; cruise missiles; fighter aircraft; intelligence, surveillance, and reconnaissance assets; precision-guided munitions; satellites; and submarines. We find strong support for our inverted-U argument. States that have rivals with moderate reconnaissance strike capabilities have over double the reconnaissance strike capabilities themselves than states that have rivals with very low or very high capabilities. Our findings hold for broader measures of the complex that closely proxy a state’s general military capabilities, narrower measures of the complex, and alternative measures of general military sophistication, indicating our theory has broad applicability. This article explains why some states invest heavily in conventional capabilities despite an already-large lead over their adversaries, and why other states instead opt to invest in alternatives rather than balancing symmetrically.

Ship detection in reefs and deep-sea with medium-high resolution images

ABSTRACT Accurate and efficient ship detection is crucial for ocean monitoring and management, especially in reefs and deep-sea, where fishing and illegal activities threaten sustainability of ecosystems. Obtaining the size of ships in reefs and deep-sea helps to identify ship types and to assess the impact of ships on marine ecosystems quantitatively. Ship detections were mainly applied at coast or river of the inland region using high spatial resolution remote sensing data due to its rich details and frequent coverage. However, the ships located in reefs and deep-sea regions were rarely studied because of price, and availability of high spatial resolution remote sensing data. Global covered medium-high resolution (10–30 m) remote sensing data (such as Sentinel-2) makes it possible to obtain the spatial-temporal distribution of ships, especially small ships. This study developed a deep learning ship detection algorithm – an enhanced Rotated-Ship Detector (RShipDet) to detect ships in reefs and deep-sea regions. RShipDet was applied to Nansha Islands based on two medium-high resolution ship detection datasets (Sentinel2-Ship and SDGSAT-Ship). These two datasets include various backgrounds of reefs and deep-sea, and complex scenarios such as cloud cover and parallel ships, adapting RShipDet to ship detection under complex circumstances in reefs and deep-sea. The results showed that: (1) Small gains: gains of 3.3% and 8.7% Average-Precision (AP) compared to Rotation-equivariant Detector (ReDet) and Faster-RCNN on Sentinel2-Ship dataset; (2) Strong generalization capabilities: 77.4% AP on SDGSAT-Ship dataset; (3) Better performance under complex conditions: RShipDet obtained more accurate ship detection results over regions with cloud cover, islands and reefs, deep-sea, and ports compared to other classical detectors. Our algorithm could be applied for better management of ocean resources and activities in reefs and deep-sea.

A fast and adaptive detection framework for genome-wide chromatin loop mapping from Hi-C data.

Chromatin loop identification plays an important role in molecular biology and 3D genomics research, as it constitutes a fundamental process in transcription and gene regulation. Such precise chromatin structures can be identified across genome-wide interaction matrices via Hi-C data analysis, which is essential for unraveling the intricacies of transcriptional regulation. Given the increasing number of genome-wide contact maps, derived from both in situ Hi-C and single-cell Hi-C experiments, there is a pressing need for efficient and resilient algorithms capable of processing data from diverse experiments rapidly and adaptively. Here, we propose YOLOOP, a novel detection-based framework that is different from the conventional paradigm. YOLOOP stands out for its speed, surpassing the performance of previous state-of-the-art (SOTA) chromatin loop detection methods. It achieves a 30-fold acceleration compared with classification-based methods, up to 20-fold acceleration compared with the SOTA kernel-based framework, and a fivefold acceleration compared with statistical algorithms. Furthermore, the proposed framework is capable of generalizing across various cell types, multiresolution Hi-C maps, and diverse experimental protocols. Compared with the existing paradigms, YOLOOP shows up to a 10% increase in recall and a 15% increase in F1-score, particularly noteworthy in the GM12878 cell line. YOLOOP also offers fast adaptability with straightforward fine-tuning, making it readily applicable to extremely sparse single-cell Hi-C contact maps. It maintains its exceptional speed, completing genome-wide detection at a 10 kb resolution for a single-cell contact map within 1 min and for a 900-cell-superimposed contact map within 3 min, enabling fast analysis of large-scale single-cell data.

Path Planning in Complex Environments Using Attention-Based Deep Deterministic Policy Gradient

The traditional Deep Deterministic Policy Gradient (DDPG) algorithm frequently exhibits a notable reduction in success rate when transferred to new environments after being trained in complex simulation settings. To address these issues, this paper adopts a Multi-Environment (Multi-Env) parallel training approach and integrates Multi-Head Attention (MHA) and Prioritized Experience Replay (PER) into the DDPG framework, optimizing the reward function to form the MAP-DDPG algorithm. This approach enhances the algorithm’s generalization capability and execution efficiency. Through comparative training and testing of the DDPG and MAP-DDPG algorithms in both simulation and real-world environments, the experimental results demonstrate that MAP-DDPG significantly improves generalization and execution efficiency over the DDPG algorithm. Specifically, in simulation environment tests, the MAP-DDPG algorithm achieved an average 30% increase in success rate and reduced the average time to reach the target point by 23.7 s compared to the DDPG algorithm. These results indicate that the MAP-DDPG algorithm significantly enhances path planning generalization and execution efficiency, providing a more effective solution for path planning in complex environments.

General Medicine and Negative Capability: Insights from the Negative Capability Conference

General Medicine and Negative Capability: Insights from the Negative Capability Conference

Accurate segmentation of intracellular organelle networks using low-level features and topological self-similarity.

Intracellular organelle networks (IONs) such as the endoplasmic reticulum (ER) network and the mitochondrial (MITO) network serve crucial physiological functions. The morphology of these networks plays a critical role in mediating their functions. Accurate image segmentation is required for analyzing the morphology and topology of these networks for applications such as molecular mechanism analysis and drug target screening. So far, however, progress has been hindered by their structural complexity and density. In this study, we first establish a rigorous performance baseline for accurate segmentation of these organelle networks from fluorescence microscopy images by optimizing a baseline U-Net model. We then develop the multi-resolution encoder (MRE) and the hierarchical fusion loss (Lhf) based on two inductive components, namely low-level features and topological self-similarity, to assist the model in better adapting to the task of segmenting IONs. Empowered by MRE and Lhf, both U-Net and Pyramid Vision Transformer (PVT) outperform competing state-of-the-art models such as U-Net++, HR-Net, nnU-Net, and TransUNet on custom datasets of the ER network and the MITO network, as well as on public datasets of another biological network, the retinal blood vessel network. In addition, integrating MRE and Lhf with models such as HR-Net and TransUNet also enhances their segmentation performance. These experimental results confirm the generalization capability and potential of our approach. Furthermore, accurate segmentation of the ER network enables analysis that provides novel insights into its dynamic morphological and topological properties. Code and data are openly accessible at https://github.com/cbmi-group/MRE.

Optimizing MFCC parameters for the automatic detection of respiratory diseases

Voice signals originating from the respiratory tract are utilized as valuable acoustic biomarkers for the diagnosis and assessment of respiratory diseases. Among the employed acoustic features, Mel Frequency Cepstral Coefficients (MFCC) are widely used for automatic analysis, with MFCC extraction commonly relying on default parameters. However, no comprehensive study has systematically investigated the impact of MFCC extraction parameters on respiratory disease diagnosis. In this study, we address this gap by examining the effects of key parameters, namely the number of coefficients, frame length, and hop length between frames, on respiratory condition examination. Our investigation uses four datasets: the Cambridge COVID-19 Sound database, the Coswara dataset, the Saarbrücken Voice Disorders (SVD) database, and a TACTICAS dataset. The Support Vector Machine (SVM) is employed as the classifier, given its widespread adoption and efficacy. Our findings indicate that the accuracy of MFCC decreases as hop length increases, and the optimal number of coefficients is observed to be approximately 30. The performance of MFCC varies with frame length across the datasets: for the COVID-19 datasets (Cambridge COVID-19 Sound database and Coswara dataset), performance declines with longer frame lengths, while for the SVD dataset, performance improves with increasing frame length (from 50 ms to 500 ms). Furthermore, we investigate the optimized combination of these parameters and observe substantial enhancements in accuracy. Compared to the worst combination, the SVM model achieves an accuracy of 81.1%, 80.6%, and 71.7%, with improvements of 19.6%, 16.10%, and 14.90% for the Cambridge COVID-19 Sound database, the Coswara dataset, and the SVD dataset respectively. To validate the generalization of these findings, we employ the Long Short-Term Memory (LSTM) model as a validation model. Remarkably, the LSTM model also demonstrates improved accuracy of 14.12%, 10.10%, and 6.68% across the datasets when utilizing the optimal combination of parameters. The optimal parameters are validated using an external voice pathology dataset (TACTICAS dataset). The results demonstrate the generalization capabilities of the optimized parameters across various pathologies, machine-learning models, and languages.

Transfer learning for well logging formation evaluation using similarity weights

Machine learning has been widely applied in well logging formation evaluation studies. However, several challenges negatively impacted the generalization capabilities of machine learning models in practical implementations, such as the mismatch of data domain between training and testing datasets, imbalances among sample categories, and inadequate representation of data model. These issues have led to substantial insufficient identification for reservoir and significant deviations in subsequent evaluations. To improve the transferability of machine learning models within limited sample sets, this study proposes a weight transfer learning framework based on the similarity of the labels. The similarity weighting method includes both hard weights and soft weights. By evaluating the similarity between test and training sets of logging data, the similarity results are used to estimate the weights of training samples, thereby optimizing the model learning process. We develop a double experts' network and a bidirectional gated neural network based on hierarchical attention and multi-head attention (BiGRU-MHSA) for well logs reconstruction and lithofacies classification tasks. Oil field data results for the shale strata in the Gulong area of the Songliao Basin of China indicate that the double experts’ network model performs well in curve reconstruction tasks. However, it may not be effective in lithofacies classification tasks, while BiGRU-MHSA performs well in that area. In the study of constructing large-scale well logging processing and formation interpretation models, it is maybe more beneficial by employing different expert models for combined evaluations. In addition, although the improvement is limited, hard or soft weighting methods is better than unweighted (i.e., average-weighted) in significantly different adjacent wells. The code and data are open and available for subsequent studies on other lithofacies layers.

A Complex Background SAR Ship Target Detection Method Based on Fusion Tensor and Cross-Domain Adversarial Learning

Synthetic Aperture Radar (SAR) ship target detection has been extensively researched. However, most methods use the same dataset division for both training and validation. In practical applications, it is often necessary to quickly adapt to new loads, new modes, and new data to detect targets effectively. This presents a cross-domain detection problem that requires further study. This paper proposes a method for detecting SAR ships in complex backgrounds using fusion tensor and cross-domain adversarial learning. The method is designed to address the cross-domain detection problem of SAR ships with large differences between the training and test sets. Specifically, it can be used for the cross-domain detection task from the fully polarised medium-resolution ship dataset (source domain) to the high-resolution single-polarised dataset (target domain). This method proposes a channel fusion module (CFM) based on the YOLOV5s model. The CFM utilises the correlation between polarised channel images during training to enrich the feature information of single-polarised images extracted by the model during inference. This article proposes a module called the cross-domain adversarial learning module (CALM) to reduce overfitting and achieve adaptation between domains. Additionally, this paper introduces the anti-interference head (AIH) which decouples the detection head to reduce the conflict of classification and localisation problems. This improves the anti-interference and generalisation ability in complex backgrounds. This paper conducts cross-domain experiments using the constructed medium-resolution SAR full polarisation dataset (SFPD) as the source domain and the high-resolution single-polarised ship detection dataset (HRSID) as the target domain. Compared to the best-performing YOLOV8s model among typical mainstream models, this model improves precision by 4.9%, recall by 3.3%, AP by 2.4%, and F1 by 3.9%. This verifies the effectiveness of the method and provides a useful reference for improving cross-domain learning and model generalisation capability in the field of target detection.

General framework for unsteady aerodynamic prediction of airfoils based on deep transfer learning

Analyzing the unsteady aerodynamic performance of airfoils under dynamic stall using computational fluid dynamics (CFD) is computationally intensive. Although deep learning models can quickly predict aerodynamic parameters, their generalization capability on a small-scale dataset is often poor. This paper presents a novel deep transfer learning (TL) framework that combines model-based and instance-based transfer methods, termed synergistic instance-model TL. This framework facilitates rapid predictions of unsteady aerodynamic performance for various airfoils and pitch oscillations from the small-scale dataset. The framework integrates the accelerated training speed of model-based methods with the dynamic dataset expansion benefits of instance-based approaches. Initially, a pre-trained Wasserstein-deep convolutional generative adversarial network (W-DCGAN) is developed, combining a convolutional neural network with a generative adversarial network to predict aerodynamic hysteresis loops for the SC1095 airfoil in the source domain. The framework then fine-tunes the pre-trained model and incorporates weighted source domain dataset into the small-scale target domain dataset, producing the transferred model W-DCGAN-TL. This approach significantly reduces prediction inaccuracies compared to model-based and non-TL methods when applied to the small-scale dataset. The framework's flexibility allows the use of pre-trained models and datasets from related aerodynamic problems to address issues with insufficient data. Consequently, it is expected to reduce the dependency on extensive datasets, enhance design efficiency, and minimize resource requirements.

A new deepfake detection model for responding to perception attacks in embodied artificial intelligence

Embodied artificial intelligence (AI) represents a new generation of robotics technology combined with artificial intelligence, and it is at the forefront of current research. To reduce the impact of deepfake technology on embodied perception and enhance the security and reliability of embodied AI, this paper proposes a novel deepfake detection model with a new Balanced Contrastive Learning strategy, named BCL. By integrating unsupervised contrastive learning and supervised contrastive learning with deepfake detection, the model effectively extracts the underlying features of fake images from both individual level and category level, thereby leading to more discriminative features. In addition, a Multi-scale Attention Interaction module (MAI) is proposed to enrich the representative ability of features. By cross-fusing the semantic features of different receptive fields of the encoder, more effective deepfake traces can be mined. Finally, extensive experiments demonstrate that our method has good performance and generalization capabilities across intra-dataset, cross-dataset and cross-manipulation scenarios.

Application of mask R-CNN for building detection in UAV remote sensing images

This study aims to tackle the challenges of low accuracy in building feature extraction and insufficient details in three-dimensional (3D) modeling faced by traditional methods, particularly in complex backgrounds. To address these issues, a method for building feature extraction based on Mask Region-Convolutional Neural Network (Mask R-CNN) is proposed. This approach combines deep learning techniques with aerial images to ensure precise and efficient automatic detection and feature extraction. Urban building images are captured through aerial photography, and building outlines are annotated to create a comprehensive dataset of building features. The Mask R-CNN-based method efficiently processes and classifies the features of the dataset, generating candidate regions for further analysis. Additionally, this method demonstrates significant advantages in building feature extraction by employing the Mask R-CNN model to generate adaptive features. Comparative analysis with models such as Convolutional Neural Network (CNN), Region-based Convolutional Neural Network (R-CNN), Fast Region-based Convolutional Neural Network (Fast R-CNN), Faster Region-based Convolutional Neural Network (Faster R-CNN), and Generative Adversarial Network (GAN) indicates that Mask R-CNN exhibits superior performance in building feature extraction. The Mask R-CNN-based approach achieved approximately 95 % classification accuracy, while also showcasing strong stability and generalization capabilities. This study provides new methodologies and insights for enhancing feature extraction in aerial building imagery, offering significant reference value for the fields of architectural design and urban planning.