Affiliation:
1. CCDC Drilling & Production Technology Research Institute
2. Institute of Modern Optics, Nankai University
3. University of Southern California
4. Tsinghua University
5. China Academy of Space Technology
6. University of Louisiana at Lafayette
Abstract
In industries such as manufacturing and safety monitoring, accurately identifying the shape characteristics of multi-opening objects is essential for the assembly, maintenance, and fault diagnosis of machinery components. Compared to traditional contact sensing methods, image-based feature recognition technology offers non-destructive assessment and greater efficiency, holding significant practical value in these fields. Although convolutional neural networks (CNNs) have achieved remarkable success in image classification and feature recognition tasks, they still face challenges in dealing with subtle features in complex backgrounds, especially for objects with similar openings, where minute angle differences are critical. To improve the identification accuracy and speed, this study introduces an efficient CNN model, ADSA-Net, which utilizes the additive self-attention mechanism. When coupled with an active light source system, ADSA-Net enables non-contact, high-precision recognition of shape features in 14 classes of rotationally symmetric objects with multiple openings. Experimental results demonstrate that ADSA-Net achieves accuracies of 100%, ≥98.04%, and ≥98.98% in identifying the number of openings, wedge angles, and opening orientations of all objects, respectively with a resolution of 1°. By adopting linear layers to replace the traditional quadratic matrix multiplication operations for key-value interactions, ADSA-Net significantly enhances computational efficiency and identification accuracy.
Funder
Shaanxi Province Innovation Talent Promotion Program-Science and Technology Innovation Team
Natural Science Foundation of Shanxi Province