Lightweight defect detection on mica sheets via focused distillation for industrial measurement
TL;DR
This paper builds a high-precision automated vision acquisition system and builds a mica sheet defect detector using Faster Region-based Convolutional Neural Network (Faster R-CNN), which achieves accurate classification and localization of defects after training on the custom dataset.
Abstract
To address the dual demands of accuracy and efficiency in detecting irregular surface defects on mica sheets, this paper proposes a comprehensive solution that integrates the design of a vision system and a lightweight detection method. First, we build a high-precision automated vision acquisition system to capture mica sheet images in batches and construct a dataset based on the defect characteristics. Next, we build a mica sheet defect detector using Faster Region-based Convolutional Neural Network (Faster R-CNN), which achieves accurate classification and localization of defects after training on the custom dataset. Finally, we introduce a knowledge distillation framework based on focused features and responses, transferring the feature extraction and response evaluation capabilities from a high-accuracy, large-capacity model to a lightweight model. Extensive experimental results demonstrate the effectiveness of the proposed method. The resulting lightweight model achieves 30.8% mean Average Precision (mAP) with only 28.1M parameters, striking a favorable balance between detection accuracy and efficiency. This provides reliable support for the insulation safety risk assessment of mica sheets.