Metal surface defect detection based on improved YOLOv5

Scientific Reports, Journal Year: 2023, Volume and Issue: 13(1)

Published: Nov. 27, 2023

Language: Английский

---

Multi-Objective Location and Mapping Based on Deep Learning and Visual Slam

Sensors, Journal Year: 2022, Volume and Issue: 22(19), P. 7576 - 7576

Published: Oct. 6, 2022

Simultaneous localization and mapping (SLAM) technology can be used to locate build maps in unknown environments, but the constructed often suffer from poor readability interactivity, primary secondary information map cannot accurately grasped. For intelligent robots interact meaningful ways with their environment, they must understand both geometric semantic properties of scene surrounding them. Our proposed method not only reduce absolute positional errors (APE) improve positioning performance system also construct object-oriented dense point cloud output model each object reconstruct indoor scene. In fact, eight categories objects are for detection using coco weights our experiments, most actual reconstructed theory. Experiments show that number points is significantly reduced. The average error Technical University Munich (TUM) datasets very small. camera reduced introduction constraints, improved. At same time, algorithm segment environment high accuracy.

Language: Английский

---

Method of recognizing sleep postures based on air pressure sensor and convolutional neural network: For an air spring mattress

Engineering Applications of Artificial Intelligence, Journal Year: 2023, Volume and Issue: 121, P. 106009 - 106009

Published: Feb. 26, 2023

Language: Английский

---

Multi-scale feature retention and aggregation for colorectal cancer diagnosis using gastrointestinal images

Engineering Applications of Artificial Intelligence, Journal Year: 2023, Volume and Issue: 125, P. 106749 - 106749

Published: July 14, 2023

Colonoscopy is considered the gold standard for colorectal cancer diagnosis and prognosis. However, existing methods are less accurate prone to overlooking lesions during gastrointestinal endoscopic examinations. Computer-assisted combined with robot-assisted minimally invasive surgery (RMIS) can significantly help medical practitioners detect treat lesions. Therefore, two novel architectures developed polyp surgical instrument segmentation aid diagnosis, assessment, treatment. Colorectal network (CCS-Net) base used in this study. It uses maximum convolutional layers near input image effective feature extraction from low-level information. In addition, CCS-Net an efficient upsampling unit efficiently increase spatial features' map size. Hence, capable of providing a fair performance satisfactory computational efficiency The multi-scale retention aggregation (MFRA-Net) final MFRA-Net improve accuracy further as it retain features transfers them deep stages network. also combines high-strided information high-level boost performance. Finally, all transferred early aggregated levels This mechanism enables maintain better compared other even challenging blur, specular reflection, low contrast, high variation cases. We evaluated both on four datasets: Kvasir-SEG, CVC-ClinicDB, Kvasir-Instrument, UW-Sinus-Surgery-Live dataset. proposed method achieves dice similarity coefficients 95.98%, 94.19%, 92.81%, 88.57% datasets. superior state-of-the-art requires only 4.9 million trainable parameters complete training. networks effectively assist health professionals procedures through instruments segmentation, respectively.

Language: Английский

---

Continuous dynamic gesture recognition using surface EMG signals based on blockchain-enabled internet of medical things

Information Sciences, Journal Year: 2023, Volume and Issue: 646, P. 119409 - 119409

Published: July 22, 2023

Language: Английский

---

Multi-Scale Feature Map Aggregation and Supervised Domain Adaptation of Fully Convolutional Networks for Urban Building Footprint Extraction

Remote Sensing, Journal Year: 2023, Volume and Issue: 15(2), P. 488 - 488

Published: Jan. 13, 2023

Automated building footprint extraction requires the Deep Learning (DL)-based semantic segmentation of high-resolution Earth observation images. Fully convolutional networks (FCNs) such as U-Net and ResUNET are widely used for segmentation. The evolving FCNs suffer from inadequate use multi-scale feature maps in their backbone neural (CNNs). Furthermore, DL methods not robust cross-domain settings due to domain-shift problems. Two scale-robust novel networks, namely MSA-UNET MSA-ResUNET, developed this study by aggregating with partial concepts pyramid network (FPN). supervised domain adaptation is investigated minimise effects between two datasets. datasets include benchmark WHU Building dataset a 5× fewer samples, 4× lower spatial resolution complex high-rise buildings skyscrapers. newly compared six state-of-the-art using five metrics: pixel accuracy, adjusted F1 score, intersection over union (IoU), Matthews Correlation Coefficient (MCC). proposed outperform majority accuracy measures both Compared larger dataset, trained on smaller one shows significantly higher robustness terms (by 18%), score 31%), IoU 27%), MCC 29%) during validation MSA-UNET. MSA-ResUNET similar improvements, concluding that when increase different complexity.

Language: Английский

---

Automatic detection of carbon steel corrosion degree based on image identification

Computational Materials Science, Journal Year: 2023, Volume and Issue: 233, P. 112717 - 112717

Published: Dec. 6, 2023

Language: Английский

---

Improved YOLOv8 algorithms for small object detection in aerial imagery

Journal of King Saud University - Computer and Information Sciences, Journal Year: 2024, Volume and Issue: 36(6), P. 102113 - 102113

Published: June 28, 2024

In drone aerial target detection tasks, a high proportion of small targets and complex backgrounds often lead to false positives missed detections, resulting in low accuracy. To improve the accuracy targets, this study proposes two improved models based on YOLOv8s, named IMCMD_YOLOv8_small IMCMD_YOLOv8_large. Each model accommodates different application scenarios. First, network structure was optimized by removing backbone P5 layer used detect large merging P4, P3, P2 layers, which are better suited for detecting medium targets; P3 serve as heads focus more targets. Subsequently, coordinate attention mechanism is integrated into backbone's C2f, create C2f_CA module that enhances model' s key information secures richer flow gradient information. multiscale feature fusion designed merge shallow deep features. Finally, Dynamic Head introduced unify perception scale, space, further enhancing capability Experimental results VisDrone2019 dataset demonstrated that, compared with achieved improvements 7.7% 5.1% [email protected] protected]:0.95, respectively, 73.0% reduction parameter count. The IMCMD_YOLOv8_large showed even significant these metrics, reaching 10.8% 7.3%, 47.7% count, displaying superior performance tasks. not only enhanced but also lightweighting, thereby proving effectiveness improvement strategies showcasing other classic models.

Language: Английский

---

Target Detection Based on Two-Stream Convolution Neural Network With Self-Powered Sensors Information

IEEE Sensors Journal, Journal Year: 2022, Volume and Issue: 23(18), P. 20681 - 20690

Published: Nov. 16, 2022

With the rapid development of artificial intelligence, a neural network is widely used in various fields. The target detection algorithm mainly based on network, but accuracy greatly related to complexity scene and texture. A RGB-D image from perspective lightweight model integration depth map overcome weak environmental illumination with self-powered sensors information proposed. This article analyzes structure YOLOv4 MobileNet, compares variation parameter numbers between depthwise separable convolution convolutional networks, combines advantages MobileNetv3 network. main three effective feature layers replaced by for initial layer extraction strengthen At same time, standard models are convolution. proposed method compared YOLOv4-MobileNetv3 this article, experimental results show that retains its original accuracy, size about 23% model, processing speed 42% higher than can still reach 83% environment poor lighting conditions.

Language: Английский

---

Fast Recognition Method for Multiple Apple Targets in Complex Occlusion Environment Based on Improved YOLOv5

Journal of Sensors, Journal Year: 2023, Volume and Issue: 2023, P. 1 - 13

Published: Feb. 6, 2023

The mechanization and intelligentization of the production process are main trends in research development agricultural products. realization an unmanned automated picking is also one hotspots China’s product engineering technology field recent years. apple-picking robot directly related to imaging research, its key use algorithms realize apple identification positioning. Aiming at problem false detection missed densely occluded targets small by robots under different lighting conditions, two recognition selected based on shape features study traditional machine learning algorithm: histogram oriented gradients + support vector (HOG SVM) a fast method for multiple complex occlusion environment improved You-Only-Look-Once-v5 (YOLOv5). first improvement CSP structure network. Using parameter reconstruction, convolutional layer (Conv) batch normalization (BN) CBL (Conv BN Leaky_relu activation function) module fused into batch-normalized Conv_B. Subsequently, CA (coordinate attention) mechanism embedded network layers designed backbone enhance expressive ability better extract targets. Finally, some with overlapping occlusions, loss function fine-tuned improve model’s recognize By comparing effects HOG SVM, Faster RCNN, YOLOv6, baseline YOLOv5 test set scenarios, $F1$ value this was increased 13.47%, 6.01%, 1.26%, 3.63%, respectively, id="M2"> 19.36%, 13.07%, 1.61%, 4.27%, illumination angles. average image time 0.27 s faster than that 0.229 0.006 YOLOv6. expected provide theoretical basis choose pertinent algorithm during operation.

Language: Английский

---