Generative adversarial network based on domain adaptation for crack segmentation in shadow environments

Computer-Aided Civil and Infrastructure Engineering, Journal Year: 2025, Volume and Issue: unknown

Published: March 2, 2025

Abstract Precision segmentation of cracks is important in industrial non‐destructive testing, but the presence shadows actual environment can interfere with results cracks. To solve this problem, study proposes a two‐stage domain adaptation framework called GAN‐DANet for crack shadowed environments. In first stage, CrackGAN uses adversarial learning to merge features from shadow‐free and datasets, creating new dataset more domain‐invariant features. second CrackSeg network innovatively integrates enhanced Laplacian filtering (ELF) into high‐resolution net enhance edges texture while out shadow information. model, addresses shift by generating features, avoiding direct feature alignment between source target domains. The ELF module effectively enhances suppresses interference, improving model's robustness Experiments show that improves accuracy, mean intersection over union value increasing 57.87 75.03, which surpasses performance existing state‐of‐the‐art algorithms.

Language: Английский

---

Frequency-Assisted Local Attention in Lower Layers of Visual Transformers

International Journal of Neural Systems, Journal Year: 2025, Volume and Issue: 35(04)

Published: Jan. 3, 2025

Since vision transformers excel at establishing global relationships between features, they play an important role in current tasks. However, the attention mechanism restricts capture of local making convolutional assistance necessary. This paper indicates that transformer-based models can attend to information without using blocks, similar kernels, by employing a special initialization method. Therefore, this proposes novel hybrid multi-scale model called Frequency-Assisted Local Attention Transformer (FALAT). FALAT introduces Window-based Positional Self-Attention (FWPSA) module limits distance query tokens, enabling contents early stage. The from value tokens frequency domain enhances diversity during self-attention computation. Additionally, traditional method is replaced with depth-wise separable convolution downsample spatial reduction for long-distance later stages. Experimental results demonstrate FALAT-S achieves 83.0% accuracy on IN-1k input size [Formula: see text] 29.9[Formula: text]M parameters and 5.6[Formula: text]G FLOPs. outperforms Next-ViT-S 0.9[Formula: text]APb/0.8[Formula: text]APm Mask-R-CNN COCO surpasses recent FastViT-SA36 3.1% mIoU FPN ADE20k.

Language: Английский

---

Semi‐supervised pipe video temporal defect interval localization

Computer-Aided Civil and Infrastructure Engineering, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 9, 2025

Abstract In sewer pipe closed‐circuit television inspection, accurate temporal defect localization is essential for effective assessment. Industry standards typically do not require time interval annotations, which are more informative but lead to additional costs fully supervised methods. Additionally, differences in scene types and camera motion patterns between inspections action (TAL) hinder the transfer of point‐supervised TAL Therefore, this study presents a semi‐supervised multi‐prototype‐based method incorporating visual odometry enhanced attention guidance (PipeSPO). The effectively leverages both unlabeled data time‐point enhances performance reduces annotation costs. Meanwhile, features exploit camera's unique videos, offering insights inform model. Experiments on real‐world datasets demonstrate that PipeSPO achieves 41.89% AP across intersection over union thresholds 0.1–0.7, improving by 8.14% current state‐of‐the‐art

Language: Английский

---

Deep line segment detection for concrete pavement distress assessment

Computer-Aided Civil and Infrastructure Engineering, Journal Year: 2025, Volume and Issue: unknown

Published: March 26, 2025

Abstract This study proposes a d eep l ine s egment etection model named DLSD, for identifying four ubiquitous line segments on concrete pavements: joint, sealed bridge expansion and roadway boundary. DLSD associates category with the triple‐point representation to encode segment. Its network employs localization head classification head, attaching several auxiliary branches integrate segment shape context. A novel dual‐attention mechanism further improves classification. From experiments, structural average precision (sAP) mean sAP of class‐agnostic class‐aware detection achieve 85.0% 73.4%, respectively. The former outperforms existing best‐performed method by 2.7%, latter sets state‐of‐the‐art performance. An automated pipeline combines cracks detect corner break shattered slab pavements an accurate distress assessment, reducing error rate ratio value from 38.7% 11.5%.

Language: Английский

---

Segmentation networks for detecting overlapping screws in 3D and color images for industrial quality control

Integrated Computer-Aided Engineering, Journal Year: 2025, Volume and Issue: unknown

Published: April 1, 2025

This study explores cost-effective, real-time strategies for bin picking in industrial quality control. An anomaly detection solution was developed a screw production plant, utilizing machine vision and AI to identify overlapping screws as anomalies. Two improvements are proposed basic initially relying on laser profiler depth images. The first improvement applies Convolutional Neural Network (CNN) the profiler's output, second replaces with camera that captures color images, applying CNN its output. tested real data using YOLOv8 Mask R-CNN segmentation models. After achieving comparable results dataset, multiple synthetic datasets, simulating different scenarios, including setups mixed screws. Results demonstrated model performance represented RGB space (red, green, blue), validating cameras an appropriate alternative. Since cheaper capture images faster, they well-suited high-speed control systems, offering significant cost advantages. Code is available at: https://github.com/enmarchi/overlapping_screws_geneneration_code .

Language: Английский

---

Self‐supervised domain adaptive approach for extrapolated crack segmentation with fine‐tuned inpainting generative model

Computer-Aided Civil and Infrastructure Engineering, Journal Year: 2025, Volume and Issue: unknown

Published: May 25, 2025

Abstract The number and proportion of aging infrastructures are increasing, thereby necessitating accurate inspection to ensure safety structural stability. While computer vision deep learning have been widely applied concrete cracks, domain shift issues often result in the poor performance pretrained models at new sites. To address this, a self‐supervised adaptation method using generative artificial intelligence based on inpainting is proposed. This approach generates site‐specific crack images labels by fine‐tuning Stable Diffusion model with DreamBooth. resulting data set then used train detection neural network learning. Evaluations across two target sets eight show average F1‐score improvements 25.82% 17.83%. A comprehensive tunnel ceiling field test further demonstrates effectiveness method. By enhancing real‐world capabilities, this supports better management.

Language: Английский

---

A Context-Dependent CNN-Based Framework for Multiple Sclerosis Segmentation in MRI

International Journal of Neural Systems, Journal Year: 2024, Volume and Issue: 35(03)

Published: Dec. 13, 2024

Despite several automated strategies for identification/segmentation of Multiple Sclerosis (MS) lesions in Magnetic Resonance Imaging (MRI) being developed, they consistently fall short when compared to the performance human experts. This emphasizes unique skills and expertise professionals dealing with uncertainty resulting from vagueness variability MS, lack specificity MRI concerning inherent instabilities MRI. Physicians manage this part by relying on their radiological, clinical, anatomical experience. We have developed an framework identifying segmenting MS scans introducing a novel approach replicating diagnosis, significant advancement field. has potential revolutionize way are identified segmented, based three main concepts: (1) Modeling uncertainty; (2) Use separately trained Convolutional Neural Networks (CNNs) optimized detecting lesions, also considering context brain, ensure spatial continuity; (3) Implementing ensemble classifier combine information these CNNs. The proposed been trained, validated, tested single modality, FLuid-Attenuated Inversion Recovery (FLAIR) MSSEG benchmark public data set containing annotated seven expert radiologists one ground truth. comparison truth each raters demonstrates that it operates similarly raters. At same time, model more stability, effectiveness robustness biases than any other state-of-the-art though using just FLAIR modality.

Language: Английский

---

Architecture Knowledge Distillation for Evolutionary Generative Adversarial Network

International Journal of Neural Systems, Journal Year: 2024, Volume and Issue: 35(04)

Published: Dec. 27, 2024

Generative Adversarial Networks (GANs) are effective for image generation, but their unstable training limits broader applications. Additionally, neural architecture search (NAS) GANs with one-shot models often leads to insufficient subnet training, where subnets inherit weights from a supernet without proper optimization, further degrading performance. To address both issues, we propose Architecture Knowledge Distillation Evolutionary GAN (AKD-EGAN). AKD-EGAN operates in two stages. First, knowledge distillation (AKD) is used during efficiently optimize subnetworks and accelerate learning. Second, multi-objective evolutionary algorithm (MOEA) searches optimal architectures, ensuring efficiency by considering multiple performance metrics. This approach, combined strategy inheritance, enhances stability quality. Experiments show that surpasses state-of-the-art methods, achieving Fréchet Inception Distance (FID) of 7.91 an Score (IS) 8.97 on CIFAR-10, along competitive results STL-10 (FID: 20.32, IS: 10.06). Code will be available at https://github.com/njit-ly/AKD-EGAN.

Language: Английский

---

Self-Supervised Image Segmentation using Meta-Learning and Multi-Backbone Feature Fusion

International Journal of Neural Systems, Journal Year: 2024, Volume and Issue: unknown

Published: Dec. 27, 2024

Few-shot segmentation (FSS) aims to reduce the need for manual annotation, which is both expensive and time-consuming. While FSS enhances model generalization new concepts with only limited test samples, it still relies on a substantial amount of labeled training data base classes. To address these issues, we propose multi-backbone few shot (MBFSS) method. This self-supervised technique utilizes unsupervised saliency pseudo-labeling, allowing be trained unlabeled data. In addition, integrates features from multiple backbones (ResNet, ResNeXt, PVT v2) generate richer feature representation than single backbone. Through extensive experimentation PASCAL-5i COCO-20i, our method achieves 54.3% 25.1% one-shot segmentation, exceeding baseline methods by 13.5% 4%, respectively. These improvements significantly enhance model’s performance in real-world applications negligible labeling effort.

Language: Английский

---