SARF-Net: A Scale-Adaptive Detection Model for Hierarchical Fusion and Precise Localization in Mini-LED Inspection

Mini-light-emitting diodes (LEDs), known for their compact size and high integration, have become pivotal in the modern display industry. Accurate and efficient automated inspection of foreign mini-LEDs is essential in industrial applications to ensure reliability and quality. However, detecting defects with significant scale heterogeneity remains a formidable challenge. Current object detectors often struggle with inadequate cross-hierarchical feature fusion and limited scale awareness in localization. To address these issues, we propose a scale-adaptive representation and fusion network (SARF-Net), a novel and scale-adaptive model designed for high-precision defect detection. SARF-Net first introduces a multiscale distributed fusion neck, which adaptively aligns receptive fields and integrates hierarchical semantics through feature alignment, global fusion, and feature injection modules, enabling comprehensive perception across complex spatial scales. To enhance efficiency and preserve defect-relevant representation, we develop a cross-partial convolutions module that leverages partial convolution to selectively emphasize the most informative channels while suppressing computational redundancy. Furthermore, we incorporate a new scale center intersection over the union loss function to improve scale-sensitive optimization through pixel variance, facilitating accurate localization for defects with ambiguous boundaries. Extensive experiments on an industrial mini-LED dataset, a supplementary micro-LED dataset, and a comparable public benchmark confirm the remarkable advantages of SARF-Net in multiscale defect detection tasks. Our method achieves state-ofthe-art performance, with 98.9% mAP50 and 94.9% mAP50:95 on the mini-LED dataset, surpassing the accuracy requirements for real-world production lines.