Aug 2026· International Conference on Computer Vision and Pattern Analysis· Vol 14296, pp. 1429611 - 1429611-7· 0 citations· 17 references
Engineering
TL;DR
A lightweight YOLOv11-based foreign object detector is designed, using StarNet to reconstruct the backbone, reducing redundant parameters and computational cost, and SDIoU is introduced for bounding box regression, which improves the localization of multi-scale targets.
Abstract
In power inspection scenarios, foreign objects are usually difficult to detect because of their small size, complex surrounding backgrounds, and limited computational resources for deployment. To improve detection efficiency under these conditions, this study designs a lightweight YOLOv11-based foreign object detector. StarNet is used to reconstruct the backbone, reducing redundant parameters and computational cost. The C3k2-DSAM module is added to the neck to reinforce fine-grained object features and weaken background interference. In addition, SDIoU is introduced for bounding box regression, which improves the localization of multi-scale targets. Experiments on the self-built dataset show that the improved model obtains 86.5% mAP@0.5, 62.1% mAP@0.5:0.95, and 89.3% Precision at 19.2 GFLOPs, demonstrating competitive accuracy with low computational cost.
The YOLOv12 network is adopted as the baseline model and the ADown module is introduced to improve downsampling efficiency while maintaining lightweight performance, and the BN-CGLU is incorporated into the A2C2f module to enhance the model’s nonlinear representation capability.
Aobo Yue, Puchun Chen, Yan Yang· Journal of Real-Time Image P...· 0 citations
Insulators play a vital role in ensuring the safe and stable operation of transmission lines. This study develops IDD-YOLO, an engineering-oriented lightweight detector for UAV-based insulator inspection, with emphasis on reducing model complexity while preserving the weak visual information of localized defects. GhostConv is used to reduce redundant computation in the backbone and neck, while a GhostConv–CARAFE lightweight neck combines efficient feature transformation with content-aware upsampling to preserve fine-grained defect information. EIoU is further employed as the bounding-box regression objective during training without adding inference-time network layers. On the IDID-Plus dataset, IDD-YOLO achieves a Precision of 83.9%, a Recall of 64.2%, and an mAP@0.5 of 66.3%, while requiring 4.2 M parameters and 9.6 GFLOPs. Compared with YOLOv11s, Precision, Recall, and mAP@0.5 increase by 4.3, 4.8, and 1.5 percentage points, respectively, whereas the parameter count and GFLOPs decrease by 54.3% and 41.5%. Although mAP@0.5:0.95 decreases slightly, the results demonstrate a competitive engineering-oriented trade-off between detection sensitivity and model complexity. The current study provides model-level evidence of lightweight design; practical deployment performance on UAV-compatible embedded hardware remains to be evaluated.
Insulators and their accessories are core components ensuring the safe operation of transmission lines. Defect detection for them is confronted with challenges such as small targets, complex backgrounds and difficulty in fitting the irregular boundaries of defects, while traditional lightweight algorithms suffer from insufficient detection performance. To tackle these challenges, this work presents a lightweight detection algorithm built upon the improved YOLOv11n: the DCNv3 (Deformable Convolution v3) is embedded into the backbone network to replace the traditional convolutional layers, enhancing the ability to extract features of irregular defects; the Inner-WIoU(Inner-Weighted Intersection over Union) loss function is employed to refine the bounding box regression. All experiments are conducted on a self-built dataset with 7,972 images and 7 defect categories. The proposed method achieves a mAP@50 of 86.4%, 3.2 percentage points higher than the baseline YOLOv11n, and outperforms mainstream lightweight detectors and small-object-oriented detectors. Meanwhile, the algorithm retains the lightweight advantage and meets the needs of on-site real-time inspection. The model maintains a lightweight structure with only 2.6 M parameters and achieves an inference speed of 38 FPS on NVIDIA RTX 4080 Super, meeting real-time inspection requirements. Meanwhile, the algorithm retains the lightweight advantage and can meet the requirements of on-site real-time inspection. Since validation is limited to a self-constructed dataset, cross-scene testing and public benchmark evaluation will be conducted in future work to further verify the model’s robustness and generalization in real transmission-line inspection scenarios.
Yu-Jie Sheng, Junjie Sang, Liutao Lu· Digital Signal and Computer...· 0 citations
This paper proposes an improved lightweight YOLO11s algorithm to address the issues of excessive model complexity and poor performance when detecting small objects in road defect detection. By incorporating the StarNet architecture to reconstruct the backbone network, the method utilizes star-shaped operations to enhance fine-grained feature extraction while reducing computational overhead. It integrates the parameter-free attention mechanism SimAM, which employs three-dimensional weighted inference to enhance target responses and suppress background noise. Additionally, the Shape-IoU loss function is introduced to improve the regression accuracy of irregularly shaped objects through dynamic shape constraints. The experimental results demonstrate that the proposed algorithm attained an mAP50 of 87.02% on the RDD2022 dataset, indicating a 4.33% increase in recall. Concurrently, it achieved a 52.7% and 50.7% reduction in parameters and computational cost, respectively, thereby demonstrating a judicious balance between detection accuracy and lightweight deployment.
Rui Zhang, Yonghao Dong· Digital Signal and Computer...· 0 citations
The detection of foreign objects in key components of high-speed trains is critical for railway safety, but existing methods struggle under low-light conditions, complex backgrounds, and small objects. To address these issues, we propose an efficient detection framework integrating an adaptive brightness enhancement network (ABEN) and a lightweight train foreign object detection network (LTFD-Net). ABEN adaptively enhances images according to their illumination, improving clarity across multiple objects and backgrounds while ensuring real-time processing. LTFD-Net combines a lightweight backbone with a multi-dimensional feature enhancement module, capturing multi-scale and contextual features to accurately detect small and complex defects with minimal computational overhead. To support realistic evaluation, we introduce the high-speed train foreign object detection (HTFD) dataset with 3,904 annotated images across five key components. Experiments show that the integrated framework achieves 85.4% mean average precision (mAP) and 112 frames per second (FPS) on HTFD, surpassing state-of-the-art methods. Independently, LTFD-Net reaches 78.3% mAP and 97 FPS on NEU-DET, demonstrating preliminary generalization capability without illumination enhancement. This work provides an efficient and practical solution for real-time foreign object detection under challenging railway conditions.
Hui Lin, Junqi Li, Baolin Liu et al.· Railway Engineering Science· 0 citations
Transmission lines are typically exposed to open environments for extended periods, rendering them susceptible to foreign object intrusions that can lead to power system failures. Conventional detection methods rarely account for complex weather conditions. To address the issue of low detection accuracy under such scenarios, this study proposes an improved foreign object detection algorithm, termed EE-DETR, based on RT-DETR. Specifically, a learnable edge enhancement module is designed to mitigate the challenges posed by weak textural features and low contrast between floating foreign objects and the background. Additionally, a dual-branch attention structure is introduced into the encoder, enabling the model to effectively capture high-frequency local details while preserving global low-frequency context in complex environments involving rain, snow, and fog. Experimental results demonstrate that, in contrast to the baseline, our approach improves precision, recall, and mAP50 by 3.7%, 4.2%, and 2.8%, respectively.
Xudong Luo· International Conference on...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
