A Novel Auto-Synthesis Dataset Approach for Fitting Recognition Using Prior Series Data

Zhang, Jie; Qin, Xinyan; Jin, Lei; Jia, Bo; Li, Bo; Li, Zhaojun; Li, Huidong; Zeng, Yujie; Song, Jie

doi:10.3390/s22124364

Cited by 3 publications

(3 citation statements)

References 25 publications

(37 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Prior synthetic datasets based on prior series data are promising for training deep learning models to recognize fittings on PTLs. We have achieved some findings in the previous studies [31,52]. The YOLOX deep learning model is trained using a prior synthetic dataset including 10,800 images of 720 × 1280 pixels, and the trained model achieves 98.38% of mAP5090 on a test dataset containing 1200 real images.…”

Section: Prior Synthetic Datasetmentioning

confidence: 56%

“…The methods of obtaining the fitting dataset include field collection [29], image stitching [30], and data synthesis [31]. Field collection methods can collect real images, but they are expensive and difficult to collect complete data.…”

Section: Related Work 21 Fitting Datasetmentioning

confidence: 99%

“…CPLID [30] included 600 normal insulators and 248 defective insulators by unmanned aerial vehicle (UAV) capture, where the defective insulators were expanded using the image stitching method due to their small number. A synthetic insulator dataset [31] included 10,800 complete insulators annotated in VOC2007 format. This study provided a synthesis method for a fitting dataset, obtaining 98% of mean average precision (mAP) using the trained YOLOX model on a real test dataset.…”

Section: Related Work 21 Fitting Datasetmentioning

confidence: 99%

See 2 more Smart Citations

A Fitting Recognition Approach Combining Depth-Attention YOLOv5 and Prior Synthetic Dataset

et al. 2022

Self Cite

View full text Add to dashboard Cite

To address power transmission lines (PTLs) traveling through complex environments leading to misdetections and omissions in fitting recognition using cameras, we propose a fitting recognition approach combining depth-attention YOLOv5 and prior synthetic dataset to improve the validity of fitting recognition. First, datasets with inspection features are automatically synthesized based on prior series data, achieving better results with a smaller data volume for the deep learning model and reducing the cost of obtaining fitting datasets. Next, a unique data collection mode is proposed using a developed flying-walking power transmission line inspection robot (FPTLIR) as the acquisition platform. The obtained image data in this collection mode has obvious time-space, stability, and depth difference, fusing the two data types in the deep learning model to improve the accuracy. Finally, a depth-attention mechanism is proposed to change the attention on the images with depth information, reducing the probability of model misdetection and omission. Test field experiments results show that compared with YOLOv5, the mAP5095 (mean average precision on step size 0.05 for thresholds from 0.5 to 0.95) of our depth-attention YOLOv5 model for fitting is 68.1%, the recall is 98.3%, and the precision is 98.3%. Among them, AP, recall, and precision increased by 5.2%, 4.8%, and 4.1%, respectively. Test field experiments verify the feasibility of the depth-attention YOLOv5. Line field experiments results show that the mAP5095 of our depth-attention YOLOv5 model for fittings is 64.6%, and the mAPs of each class are improved compared with other attention mechanisms. The inference speed of depth-attention YOLOv5 is 3ms slower than the standard YOLOv5 model and 10ms to 15ms faster than other attention mechanisms, verifying the validity of the depth-attention YOLOv5. The proposed approach improves the accuracy of the fitting recognition on PTLs, providing a recognition and localization basis for the automation and intelligence of inspection robots.

show abstract

Section: Prior Synthetic Datasetmentioning

confidence: 56%