Digital Signal Processing, Год журнала: 2025, Номер unknown, С. 105343 - 105343
Опубликована: Май 1, 2025
Язык: Английский
Measurement, Год журнала: 2025, Номер unknown, С. 116813 - 116813
Опубликована: Янв. 1, 2025
Язык: Английский
Процитировано
1
Fire, Год журнала: 2025, Номер 8(4), С. 125 - 125
Опубликована: Март 24, 2025
To address the challenges of high algorithmic complexity and low accuracy in current fire detection algorithms for highway tunnel scenarios, this paper proposes a lightweight algorithm, FIRE-YOLOv8s. First, novel feature extraction module, P-C2f, is designed using partial convolution (PConv). By dynamically determining kernel’s range action, module significantly reduces model’s computational load parameter count. Additionally, ADown introduced downsampling, employing branching design to minimize requirements while preserving essential information. Secondly, neck fusion network redesigned CNN-based cross-scale (CCFF). This leverages operations achieve efficient fusion, further reducing model enhancing efficiency multi-scale features. Finally, dynamic head introduced, incorporating multiple self-attention mechanisms better capture key information complex scenes. improvement enhances robustness detecting targets under challenging conditions. Experimental results on self-constructed dataset demonstrate that, compared baseline YOLOv8s, FIRE-YOLOv8s by 47.2%, decreases number parameters 52.2%, size 50% original, achieving 4.8% 1.7% increase [email protected]. Furthermore, deployment experiments emergency firefighting robot platform validate algorithm’s practical applicability, confirming its effectiveness real-world scenarios.
Язык: Английский
Процитировано
0
Drones, Год журнала: 2025, Номер 9(4), С. 293 - 293
Опубликована: Апрель 10, 2025
In Unmanned Aerial Vehicle (UAV) target detection tasks, issues such as missing and erroneous detections frequently occur owing to the small size of targets complexity image background. To improve these issues, an improved algorithm named RLRD-YOLO, based on You Only Look Once version 8 (YOLOv8), is proposed. First, backbone network initially integrates Receptive Field Attention Convolution (RFCBAMConv) Module, which combines Convolutional Block Module (CBAM) (RFAConv). This integration improves issue shared attention weights in receptive field features. It also mechanisms across both channel spatial dimensions, enhancing capability feature extraction. Subsequently, Large-Scale Kernel (LSKA) integrated further optimize Spatial Pyramid Pooling Fast (SPPF) layer. enhancement employs a large-scale convolutional kernel capture intricate features minimize background interference. enhance fusion effectively integrate low-level details with high-level semantic information, Reparameterized Generalized Feature Network (RepGFPN) replaces original architecture neck network. Additionally, small-target layer added model’s ability perceive targets. Finally, detecting head replaced Dynamic Head, designed localization accuracy complex scenarios by optimizing for Scale Awareness, Task Awareness. The experimental results showed that RLRD-YOLO outperformed YOLOv8 VisDrone2019 dataset, achieving improvements 12.2% [email protected] 8.4% [email protected]:0.95. surpassed other widely used object methods. Furthermore, HIT-HAV dataset demonstrate sustains excellent precision infrared UAV imagery, validating its generalizability diverse scenarios. was deployed validated typical airborne platform, Jetson Nano, providing reliable technical support improvement algorithms aerial their practical applications.
Язык: Английский
Процитировано
0
Digital Signal Processing, Год журнала: 2025, Номер unknown, С. 105320 - 105320
Опубликована: Май 1, 2025
Язык: Английский
Процитировано
0
Digital Signal Processing, Год журнала: 2025, Номер unknown, С. 105343 - 105343
Опубликована: Май 1, 2025
Язык: Английский
Процитировано
0