A study on an embedded system for workpiece defect detection based on a multiscale image algorithm using deep learning

Abstract

Surface defect detection of workpieces is a core task in the field of industrial visual quality inspection. Due to differences in workpiece material, lighting conditions, and defect causes, defects exhibit significant scale differences in images. Existing detection models face a contradiction between computing resources and detection accuracy when deployed on embedded platforms. A lightweight detection network based on multi-scale feature fusion has been designed, which uses depthwise separable convolution to reduce model computation and introduces a feature pyramid structure to achieve cross scale semantic information fusion. To meet the requirements of embedded deployment, 8-bit integer quantization and channel pruning techniques are used to compress the model volume, and the inference process is optimized based on the ARM (Advanced RISC Machine) architecture instruction set. The experimental results on the NEU-DET (Northeastern University Defect Detection Dataset) dataset show that the proposed method achieves an effective balance between model parameter quantity and detection accuracy.