CRMA-Det: An RT-DETR-Based Model for Instance-Level Ordinal Maturity Detection of Chinese Hickory

Chemoecology · Springer Nature · 2026-10-26 · DOI: 10.1007/s10341-026-02066-x

Аннотация

Chinese hickory ( Carya cathayensis ) is an important woody oil crop and a source of non-wood forest products in China. In natural orchards, reliable instance-level maturity recognition supports harvest timing, planning, and fruit-level target selection. Adjacent stages may differ only in subtle pericarp color, surface patches, roughness, and cracking. We propose Cross-stage Recalibrated Maturity-Aware Detector (CRMA-Det), an end-to-end visual perception model based on real-time detection transformer (RT-DETR)-R18VD for fruit localization and four-stage ordinal maturity detection (M0–M3). It combines cross-stage partial reconstruction of PResNet Stages 3 and 4 (CSP-PResNet-S34), mean-preserving efficient channel attention at P4 (MP-ECA-P4), and cumulative ordinal auxiliary supervision. Together, these components retain weak cues, recalibrate maturity-relevant channels, and model stage order during training. We re-annotated 1661 valid images from the published CaryaData orchard corpus. Across three independent runs with $$\lambda_{\mathrm{ord}}=0.800$$, CRMA-Det achieved mean test precision of 0.783, recall of 0.827, F1 score of 0.804, mAP50 of 0.868, and mAP50:95 of 0.800. It reached 95.160 model-only frames per second (FPS) under a fixed-shape CUDA Graph protocol. Against RT-DETR-R18VD, it improved precision, recall, F1, mAP50, and mAP50:95 by 2.7, 4.6, 3.8, 2.3, and 2.0 percentage points, respectively, while reducing parameter count by 2.86% and giga floating-point operations (GFLOPs) by 5.10%. CRMA-Det therefore provides an efficient perception module for orchard maturity surveying and harvest decision support.

Читать у издателя →

Похожие статьи