Real-Time Surface Defect Detection with Compound Scaling Dynamic Neural Networks

This paper presents a compound scaling dynamic network architecture designed for real-time surface defect detection in industrial production lines, addressing the challenges posed by varying time constraints and computational resource allocation. We develope a lightweight UNet-based dynamic network using a ResNet-18 backbone, incorporating optional convolutional blocks at the end of each network stage to simultaneously scale network depth and width. To efficiently train this dynamic network, an incremental training strategy is proposed to minimize interference between different paths. The network is trained and tested on the Severstal Steel Defect Detection dataset, demonstrating that the compound scaling approach significantly improves accuracy at the same Floating Point Operations (FLOPs) cost compared to single-dimension scaling methods. Additionally, a scheduling algorithm is designed to dynamically allocate computational resources and adjust network levels, ensuring real-time requirements are met in multitask environments. Experimental results validate the robustness of this scheduling algorithm in handling both periodic and aperiodic tasks, as well as a increasing number of tasks. Under varying time constraints, our compound scaling approach shows higher defect detection accuracy by at most 4.0% and 4.6% respectively, compared to network depth-only scaling method and resolution-only scaling method, while mamtaining a quite low deadline miss rate under 0.3%.