A Domestic Trash Detection Model Based on Improved YOLOX

Sensors, Journal Year: 2022, Volume and Issue: 22(18), P. 6974 - 6974

Published: Sept. 15, 2022

Domestic trash detection is an essential technology toward achieving a smart city. Due to the complexity and variability of urban scenarios, existing algorithms suffer from low rates high false positives, as well general problem slow speed in industrial applications. This paper proposes i-YOLOX model for domestic based on deep learning algorithms. First, large number real-life images are collected into new image dataset. Second, lightweight operator involution incorporated feature extraction structure algorithm, which allows layer establish long-distance relationships adaptively extract channel features. In addition, ability distinguish similar features strengthened by adding convolutional block attention module (CBAM) enhanced network. Finally, design residual head reduces gradient disappearance accelerates convergence loss values allowing perform better classification regression acquired layers. this study, YOLOX-S chosen baseline each enhancement experiment. The experimental results show that compared with mean average precision (mAP) improved 1.47%, parameters reduced 23.3%, FPS 40.4%. practical applications, achieves accurate recognition natural scenes, further validates generalization performance provides reference future research.

Language: Английский

---

An Image Object Detection Model Based on Mixed Attention Mechanism Optimized YOLOv5

Electronics, Journal Year: 2023, Volume and Issue: 12(7), P. 1515 - 1515

Published: March 23, 2023

As one of the more difficult problems in field computer vision, utilizing object image detection technology a complex environment includes other key technologies, such as pattern recognition, artificial intelligence, and digital processing. However, because an can be complex, changeable, highly different, easily confused with target, target is affected by factors, insufficient light, partial occlusion, background interference, etc., making multiple targets extremely robustness algorithm low. How to make full use rich spatial information deep texture accurately identify type location urgent problem solved. The emergence neural networks provides effective way for feature extraction utilization. By aiming at above problems, this paper proposes model based on mixed attention mechanism optimization YOLOv5 (MAO-YOLOv5). proposed method fuses local features global so better enrich expression ability map effectively detect objects large differences size within image. Then, added weigh each channel, enhance features, remove redundant improve recognition network towards background. results show that has higher precision faster running speed perform object-detection tasks.

Language: Английский

---

Silicon energy bulk material cargo ship detection and tracking method combining YOLOv5 and DeepSort

Energy Reports, Journal Year: 2023, Volume and Issue: 9, P. 151 - 158

Published: Jan. 30, 2023

Silicon widely contained in sand is expected to become a new energy material with environmental protection, safety and low cost. The intelligent management of the exploitation transportation such resources has an urgent demand. Aiming at solving problem detection rate silicon bulk cargo ship targets river monitoring videos, this paper proposes multi-ship tracking method based on YOLOv5x combined DeepSort algorithm. In order improve detector recognition efficiency, CIoU Loss used as target bounding box regression loss function instead GIoU speed up while improving localization accuracy; NMS replaced by DIoU-NMS tackle missed when are dense. structure appearance feature extraction network adjusted trained self-built dataset reduce identity switching caused occlusion. experimental results show that Pytorch framework makes model converge fast accurate mAP reaching 95.6%, then can achieve multiple types ships. This provide effective technical support for supervision illegal mining material.

Language: Английский

---

FPGA Implementation of a Deep Learning Acceleration Core Architecture for Image Target Detection

Applied Sciences, Journal Year: 2023, Volume and Issue: 13(7), P. 4144 - 4144

Published: March 24, 2023

Due to the flexibility and ease of deployment Field Programmable Gate Arrays (FPGA), more studies have been conducted on developing optimizing target detection algorithms based Convolutional Neural Networks (CNN) models using FPGAs. Still, these focus improving performance core algorithm hardware structure, with few focusing unified architecture design corresponding optimization techniques for model, resulting in inefficient overall model performance. The essential reason is that do not address arithmetic power, speed, resource consistency. In order solve this problem, we propose a deep learning acceleration FPGAs, which designed CNN models, multi-channel parallelization network improve scheduling tasks intensive computation pipelining meet algorithm’s data bandwidth requirements unifying speed area orchestrated matrix save resources. proposed framework achieves 14 Frames Per Second (FPS) inference TinyYolo at 5 Giga Operations (GOPS) 30% higher running clock frequency, 2–4 times 28% Digital Signal Processing (DSP) utilization efficiency less than 25% FPGA usage.

Language: Английский

---

A New Vehicle Detection Framework Based on Feature-Guided in the Road Scene

Computers, materials & continua/Computers, materials & continua (Print), Journal Year: 2024, Volume and Issue: 78(1), P. 533 - 549

Published: Jan. 1, 2024

Vehicle detection plays a crucial role in the field of autonomous driving technology. However, directly applying deep learning-based object algorithms to complex road scene images often leads subpar performance and slow inference speeds vehicle detection. Achieving balance between accuracy speed is for real-time real-world scenes. This paper proposes high-precision fast detector called feature-guided bidirectional pyramid network (FBPN). Firstly, tackle challenges like occlusion significant background interference, efficient feature filtering module (EFFM) introduced into network, which amplifies disparities features background. Secondly, proposed global attention localization (GALM) model neck effectively perceives detailed position information target, improving both model. Finally, small-scale vehicles further enhanced through utilization four-layer structure. Experimental results show that FBPN achieves an average precision 60.8% 97.8% on BDD100K KITTI datasets, respectively, with reaching 344.83 frames/s 357.14 frames/s. demonstrates its effectiveness superiority by striking speed, outperforming several state-of-the-art methods.

Language: Английский

---

Optimization Study of Coal Gangue Detection in Intelligent Coal Selection Systems Based on the Improved Yolov8n Model

Electronics, Journal Year: 2024, Volume and Issue: 13(21), P. 4155 - 4155

Published: Oct. 23, 2024

To address the low recognition accuracy of models for coal gangue images in intelligent preparation systems—especially identifying small target due to factors such as camera angle changes, illumination, and motion blur—we propose an improved separation model, Yolov8n-improvedGD(GD—Gangue Detection), based on Yolov8n. The optimization strategy includes integrating GCBlock(Global Context Block) from GCNet(Global Network) into backbone network enhance model’s ability capture long-range dependencies improve performance. CGFPN (Contextual Guidance Feature Pyramid module is designed optimize feature fusion expression capabilities. GSConv-SlimNeck architecture employed computational efficiency map capabilities, thereby improving robustness. A 160 × scale detection head incorporated sensitivity detection, mitigate effects low-quality data, localization accuracy.

Language: Английский

---

UAV measurements and AI-driven algorithms fusion for real estate good governance principles support

International Journal of Applied Earth Observation and Geoinformation, Journal Year: 2024, Volume and Issue: 134, P. 104229 - 104229

Published: Oct. 28, 2024

Language: Английский

---

SARFB: Strengthened Asymmetric Receptive Field Block for Accurate Infrared Ship Detection

IEEE Sensors Journal, Journal Year: 2023, Volume and Issue: 23(5), P. 5028 - 5044

Published: Jan. 20, 2023

Convolutional neural network (CNN)-based detection has shown great potential in accurate infrared (IR) ship detection. Typically, IR images exhibit a lack of texture details, whereas the sizes targets are extremely multiscale, making it difficult to accurately detect targets. Herein, we propose novel strengthened asymmetric receptive field block (SARFB) for The SARFB contains an (ARFB), spatial pyramid pooling (SPP) block, and skip connections. Through these components, is able fuse local global features, enriching expressive ability multiscale target Furthermore, because there no publicly available dataset detection, created single-frame (SFISD) dataset, providing first public benchmark testing performance. In comparative studies, mAP_0.5 Yolov5 with reached 93.3%, outperforming other state-of-the-art methods. Finally, performed experiments on unmanned surface vehicle (USV) equipped camera. results show superior robustness our proposed method, especially when information lacking, multiscale. SFISD at https://github.com/echoo-sky/SFISD .

Language: Английский

---

Fusing Self-Attention and CoordConv to Improve the YOLOv5s Algorithm for Infrared Weak Target Detection

Sensors, Journal Year: 2023, Volume and Issue: 23(15), P. 6755 - 6755

Published: July 28, 2023

Convolutional neural networks have achieved good results in target detection many application scenarios, but convolutional still face great challenges when facing scenarios with small sizes and complex background environments. To solve the problem of low accuracy infrared weak scenes, considering real-time requirements task, we choose YOLOv5s algorithm for improvement. We add Bottleneck Transformer structure CoordConv to network optimize model parameters improve performance network. Meanwhile, a two-dimensional Gaussian distribution is used describe importance pixel points frame, normalized Guassian Wasserstein distance (NWD) measure similarity between prediction frame true characterize loss function targets, which will help highlight targets flat positional deviation transformation accuracy. Finally, through experimental verification, compared other mainstream algorithms, improved this paper significantly improves accuracy, mAP reaching 96.7 percent, 2.2 percentage higher Yolov5s.

Language: Английский

---

Infrared Maritime Object Detection Network With Feature Enhancement and Adjacent Fusion

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Journal Year: 2024, Volume and Issue: 17, P. 5750 - 5760

Published: Jan. 1, 2024

As a crucial maritime search and rescue method, infrared object detection is critical in influencing the success rate. Research on images limited, problems of smaller sizes, more substantial noise, less detailed information still need to be solved. To tackle these problems, we proposed an network with feature enhancement adjacent fusion. A spatial module semantic are designed enhance location dim small targets deep information, respectively. We fusion fully use multi-scale information. built dataset compared method existing advanced traditional learning methods. The achieves better results.

Language: Английский

---