DFINE-AGS: A Representation-Preserving Transformer Detector for Potato Detection in Complex Field Environments
Year: 2026
Type: preprint
Abstract: Reliable recognition of potato tubers in open-field scenes remains difficult because targets frequently exhibit soil adhesion, partial burial, mutual occlusion, large apparent-scale variation, uneven illumination, and strong visual similarity to surrounding soil and residues. We propose DFINE-AGS, a D-FINE-derived detector formulated as a coordinated representation-preservation framework for these field-induced degradations. GC-BRA combines global semantic anchoring, content-dependent region routing, and local spatial compensation to improve target-background discrimination; SPDA-Conv preserves phase-structured local evidence before spatial reduction; and the P3/P4/P5 Zoom_cat-ScalSeq-ASF pathway models ordered cross-scale interactions between fine spatial detail and high-level semantics. On the self-collected potato dataset, DFINE-AGS achieved 90.4% precision, 91.5% recall, a 90.9% F1-score, and 90.0% [email protected]. TensorRT inference on an NVIDIA Jetson Orin Nano reached 37.4 FPS with 90.3% [email protected]. An independent field evaluation containing 1,297 manually verified potato targets yielded 89.67% precision and 90.36% recall. These results demonstrate that DFINE-AGS provides a robust and edge-deployable detection framework for complex field environments while retaining a favorable accuracy-efficiency trade-off.
Source: Research Square
Institutions: Qingdao Agricultural University, Hainan University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: green
OpenAlex ID: W7225912322
