Exploring Variants of Feature Pyramid Networks (FPN) for Object Detection: A Comprehensive Review

Feature Pyramid Networks (FPN) have become a standard multi-scale representation for modern visual recognition. This review consolidates core FPN design principles-top-down pathways, lateral fusion, and scale-aware heads and synthesizes recent variants such as BiFPN, PANet/PAFPN, ssFPN, Refine-FPN, and attention-guided fusion. We examine how these necks integrate with common backbones and detectors, summarize typical training setups and datasets (e.g., COCO, DOTA, SIRST), and report qualitative benefits frequently claimed in the literature, including improved small-object sensitivity, faster convergence, and better robustness to scale variance. Beyond natural images, we highlight applications in specialized domains (remote sensing, infrared, underwater), noting when patterns plausibly translate to medical image analysis. A functional comparison of representative detectors clarifies pipeline choices, loss designs, and computational trade-offs. We conclude with open challenges reliable gains across backbones, practical deployment under resource constraints, and standardized evaluation for domain transfer-and provide curated tables that connect variants to tasks, detectors, and the corresponding canonical references.

کلیدواژه ها:

Feature Pyramid Network (FPN) ، Multiscale Object Detection ، Deep Learning Architectures ، Convolutional Neural Networks (CNNs) ، Image Analysis ، Multi-scale Object Detection ، Lateral Connections ، Semantic Feature Fusion