Activation-Tunable Network for Surface Defect Detection

Xunkuai Zhou, Xi Chen, Jie Chen, Ben M. Chen

Keywords: Edge-computing device, Deep learning, Defect detection

Abstract: Vision-based defect detection effectively monitors the condition and quality of construction and industrial products. This work presents an accurate detection network augmented by an environmental interaction module and a flexible, tunable activation function. The environmental interaction module is designed to localize and detect defects more accurately, while the flexible activation improves accuracy without increasing parameters. To mitigate information loss after downsampling, we restructure features and introduce a simple deep-global fusion module that integrates deep and global cues to enhance detection performance. Extensive experiments demonstrate the superiority of the proposed network, and real-world tests highlight its portability and practicality. On an edge-computing device, the model achieves real-time inference at 15 FPS, underscoring its suitability for resource-constrained deployment. Furthermore, the proposed activation function enhances the nonlinear representational capacity of neural networks, outperforming 20 widely used activation functions in detection accuracy.

Submission Type: Novel research

Student Paper: No

Demo Or Video: No

Public Extended Abstract: Yes

Submission Number: 5