Test-time fine-tuning of large-scale vision foundation models for data-efficient electronic waste segmentation

Electronic waste (e-waste) recycling is crucial for resource conservation and environmental sustainability. Recent advances in artificial intelligence (AI), particularly deep learning, have shown strong potential to advance waste recognition and recycling. However, deep learning-based e-waste recognition remains challenging due to heterogeneous, cluttered images and limited labeled data. Model adaptation is further constrained by high computational costs and lengthy adaptation times. Large-scale vision foundation models (LVMs) provide strong generalization across visual domains, but their direct application to e-waste imaging remains limited for fine-grained segmentation. Here, we present a deep learning-based segmentation approach that integrates test-time fine-tuning with LVMs for automated e-waste recognition. The proposed approach adapts LVMs at inference time using a small, diverse, and contextually relevant support set, avoiding the need for offline domain-specific fine-tuning. To construct this support set, we introduce active retrieval , which combines cosine similarity-based preselection with an informativeness acquisition function to identify query-specific, non-redundant support images. Evaluated on a curated e-waste dataset, the method reduces end-to-end processing time by approximately sixfold compared with parameter-efficient fine-tuning, while maintaining reliable segmentation accuracy using less domain-specific data. This scalable approach supports practical AI deployment for e-waste management and improved resource recovery.