Sparsifying Instance Segmentation Models for Efficient Vision-Based Industrial Recycling

Recycling is essential to the circular economy. However, efficient material sorting, particularly in steel scrap recycling, remains challenging due to material diversity and contamination. Visual computing via deep learning offers a significant promise in automation, with models such as YOLO and Mask R-CNN excelling in object detection and segmentation. However, high computational requirements often limit industrial deployment, which necessitates more efficient algorithmic solutions targeted for such applied machine learning problems. We introduce a novel approach to prune large image segmentation models based on instance-based importance scores (IBIS), specifically tailored to the problem of instance segmentation for automated steel scrap recycling. Our method identifies and prunes low priority parameters by leveraging parameter importance scores estimated by considering the presence of recyclable instances to be segmented in the frames. Moreover, we utilize a novel custom dataset constructed for the instance segmentation task during copper and steel scrap recycling, which involves recyclable objects of different sizes with various levels of difficulty. Our evaluations demonstrate promising computational efficiency gains without significant performance drops, while also enabling powerful out-of-distribution generalization, a game-changing capability. Finally, we discuss the potential of our work for real-world industrial applications, enabling resource-efficient deep learning deployment in large-scale automated sorting systems.