Photoelastic Stress Field Recovery Using Deep Convolutional Neural Network

Frontiers in Bioengineering and Biotechnology, Journal Year: 2022, Volume and Issue: 10

Published: March 21, 2022

Recent work has shown that deep convolutional neural network is capable of solving inverse problems in computational imaging, and recovering the stress field loaded object from photoelastic fringe pattern can also be regarded as an problem process. However, formation affected by geometry specimen experimental configuration. When produces complex distribution, traditional analysis methods still face difficulty unwrapping. In this study, a based on encoder-decoder structure proposed, which accurately decode distribution information images generated under different configurations. The proposed method validated synthetic dataset, quality model evaluated using mean squared error (MSE), structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), other evaluation indexes. results show recovery achieve average performance more than 0.99 SSIM.

Language: Английский

---

Volatile threshold switching memristor: An emerging enabler in the AIoT era

Journal of Semiconductors, Journal Year: 2023, Volume and Issue: 44(5), P. 053102 - 053102

Published: May 1, 2023

Abstract With rapid advancement and deep integration of artificial intelligence the internet-of-things, things has emerged as a promising technology changing people’s daily life. Massive growth data generated from devices challenges AIoT systems information collection, storage, processing communication. In review, we introduce volatile threshold switching memristors, which can be roughly classified into three types: metallic conductive filament-based TS devices, amorphous chalcogenide-based ovonic metal-insulator transition based devices. They play important roles in high-density energy efficient computing hardware security for systems. Firstly, brief introduction is exhibited to describe categories (materials characteristics) And then, mechanisms types are discussed systematically summarized. After that, attention focused on applications 3D cross-point memory with high storage-density, neuromorphic computing, (true random number generators physical unclonable functions), others (steep subthreshold slope transistor, logic etc. ). Finally, major future outlook memristors presented.

Language: Английский

---

Domain Adaptation With Contrastive Learning for Object Detection in Satellite Imagery

IEEE Transactions on Geoscience and Remote Sensing, Journal Year: 2024, Volume and Issue: 62, P. 1 - 15

Published: Jan. 1, 2024

State-of-the-art object detection methods applied to satellite and drone imagery largely fail identify cross-domain small dense objects. The high content variability in the overhead is due different sensors, terrestrial regions, lighting conditions, image acquisition time of day. Moreover, number size objects aerial are very than consumer data. We propose a pipeline that improves feature extraction process by spatial pyramid pooling, cross-stage partial networks, heatmap-based region proposal networks. Next, we instance-aware difficulty score adapts overall focal loss improve localization identification. Finally, add two progressive domain adaptation blocks using contrastive learning pipeline. align local global features extracted from customized CSP Darknet backbone, as levels alignment alleviate degradation identification previously unseen datasets. create first-ever benchmark for task highly imbalanced datasets with significant gaps dominant existing benchmarks—the proposed method results up 7.4% 4.6% increase mAP over best state-of-art DOTA NWPU-VHR10 datasets, respectively.

Language: Английский

---

Crackwave R-convolutional neural network: A discrete wavelet transform and deep learning fusion model for underwater dam crack detection

Structural Health Monitoring, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 3, 2025

Crack detection is an essential part of structural health monitoring (SHM) for underwater dams, which crucial preventing potential failures and ensuring the long-term stability. Deep learning-based image processing algorithms have become a research hotspot in field crack detection. However, complex environment has posed challenges to dam To address these issues, we propose CrackWave R-convolutional neural network (CW R-CNN), novel model that fuses discrete wavelet transform (DWT) deep learning. The proposed introduces backbone network, DwtResNet, incorporates DWT comprehensively extract frequency-domain features from images. overcome limitations Intersection over Union (IoU), particularly when predicted ground truth bounding boxes do not overlap, employ generalized IoU (GIoU) function. Furthermore, apply soft nonmaximum suppression (NMS) algorithm reduce risk missing fine cracks. In addition, utilized self-developed acquisition robot capture large number images, forming self-acquired dataset. Evaluating on this dataset showed its MAP_0.5 outperformed SSD, YOLOv5, conventional Faster R-CNN. proved more effective than other models, especially detecting cracks handling backgrounds. These experimental results only demonstrate effectiveness CW R-CNN but also highlight application SHM. It provides technical support safe structures.

Language: Английский

---

Biomimetic Vision for Zoom Object Detection Based on Improved Vertical Grid Number YOLO Algorithm

Frontiers in Bioengineering and Biotechnology, Journal Year: 2022, Volume and Issue: 10

Published: May 20, 2022

With the development of bionic computer vision for images processing, researchers have easily obtained high-resolution zoom sensing images. The drones equipped with high-definition cameras has greatly increased sample size and image segmentation target detection are important links during process information. As biomimetic remote usually prone to blur distortion in imaging, transmission processing stages, this paper improves vertical grid number YOLO algorithm. Firstly, light shade a were abstracted, grey-level cooccurrence matrix extracted feature parameters quantitatively describe texture characteristics image. Simple Linear Iterative Clustering (SLIC) superpixel method was used achieve light/dark scenes, saliency area obtained. Secondly, model segmenting dark scenes established made dataset meet recognition standard. Due refraction passing through lens other factors, difference contour boundary value between pixel background would make it difficult detect target, pixels main part separated be sharper edge detection. Thirdly, algorithm an improved proposed real time on processed array. adjusted aspect ratio modified grids network structure by using 20 convolutional layers five maximum aggregation layers, which more accurately adapted "short coarse" identified object information density. Finally, comparison mainstream algorithms different environments, test results aid showed that high spatial resolution images, higher accuracy than had real-time performance accuracy.

Language: Английский

---