Research on Propeller Blades Sanding Degree Recognition Algorithm Based on Improved DeepLabV3+ Network

In the overhaul process of airplane propeller blades, the foam filling on the shoulder of the blades needs to be ground off. To achieve automated recognition of the sanding degree of propeller blades, an image semantic segmentation algorithm based on the improved DeepLabV3+ network is proposed. The...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEEE International Conference on Mechatronics, Robotics and Automation (Online) S. 157 - 163
Hauptverfasser: Cao, Siying, Yu, Baocheng, Xu, Wenxia
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 20.09.2024
Schlagworte:
Accuracy
Airplane Propeller blades sanding image
Airplanes
Blades
Convolution
DeepLabV3
Image recognition
Inference algorithms
Lightweight
Propellers
Robotics and automation
Semantic segmentation
Training
ISSN:2996-380X
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In the overhaul process of airplane propeller blades, the foam filling on the shoulder of the blades needs to be ground off. To achieve automated recognition of the sanding degree of propeller blades, an image semantic segmentation algorithm based on the improved DeepLabV3+ network is proposed. The algorithm adopts a lightweight network, MobileNetV2, to replace the original backbone network of DeepLabv3+, reducing computational complexity. The atrous spatial pyramid pooling (ASPP) feature extraction module is improved by replacing the atrous convolution with depthwise separable convolution and introducing the Convolutional Block Attention Module(CBAM), enhancing the overall training efficiency and segmentation accuracy of the model. Group normalization (GN) is used instead of batch normalization (BN) to further improve model performance. The experimental results on the dataset collected from the propeller blades sanding site show that the proposed algorithm achieves a mean intersection over union (MioU) of 91.40%, a mean pixel accuracy (MPA) of 95.63%, and an inference speed (FPS) of 33.88. Compared to the current mainstream segmentation algorithms such as DeepLabV3+, PSPNet, and HRNet, the improved algorithm demonstrates enhanced segmentation accuracy and inference speed.
ISSN:2996-380X
DOI:10.1109/ICMRA62519.2024.10809026