关键词: 香石竹分级, 目标检测, Flor-YOLO, 轻量化, 小波池化, 频域分析

Abstract:

[Objective] Carnation (Dianthus caryophyllus L.) is one of the most economically valuable cut flower crops worldwide. Postharvest openness is a key quality indicator influencing pricing, logistics tolerance, and shelf life. However, manual grading is inefficient and subjective due to dense petal overlap and complex edge structures. With the shift toward large-scale production and rising labor costs, accurate automated grading has become essential. Existing object detection models face a trade-off between computational efficiency and feature fidelity: High-precision architectures are computationally expensive for edge deployment, while lightweight models often lack sufficient feature representation. Additionally, conventional spatial downsampling introduces spectral aliasing, leading to the loss of high-frequency petal texture information and limiting the separability of adjacent openness grades. Therefore, a lightweight yet detail-preserving detection framework is required. To address this need, Flor-YOLO (Flower openness recognition You Only Look Once) is proposed integrating frequency-domain perception with structural re-parameterization for efficient and accurate carnation openness grading. [Methods] Based on the YOLO11n baseline, the Flor-YOLO architecture was proposed with targeted improvements to the backbone, downsampling mechanism, and detection head. Backbone reconstruction: A lightweight LiteChimeraNet was constructed to enhance feature expression under limited computing power. A RepStem re-parameterization module was introduced at the input stage to establish an anti-aliasing mechanism via multi-branch training and single-path inference. Simultaneously, the C3k2_PConv module, utilizing partial convolution, was integrated to reduce memory access cost (MAC) and focus computation on petal foregrounds. Additionally, a RepNCSPELAN4_CAA module embedded with context anchor attention was incorporated in deep layers to capture long-range dependencies of the global flower topology. Frequency-domain downsampling: To mitigate texture aliasing and detail loss caused by spatial downsampling, a WaveletPool module was introduced. Utilizing the 2D discrete wavelet transform (2D-DWT), this module orthogonally decomposed feature maps into low- and high-frequency sub-bands, explicitly preserving high-frequency information in horizontal, vertical, and diagonal directions to alleviate spectral aliasing. Detection head optimization: A lightweight shared detail-enhanced detection head (SDL-Head) was designed. It reduced parameter redundancy through cross-scale weight sharing and incorporated detail-enhanced convolution (DEConv), fusing central and angular difference operators, to boost sensitivity to the geometric morphology of petal edges. Furthermore, a scale-adaptive layer combined with Intersection over Union (IoU)-aware soft labels was applied to improve multi-scale feature alignment. A dataset comprising 1 748 original images of "Red kang" carnations was collected and expanded to 6 580 samples via hybrid data augmentation. The model was trained on an NVIDIA RTX 4060 GPU for 250 epochs using SGD optimization, and comparative evaluations were conducted against the YOLO series, NanoDet-m, and Hyper-YOLO-t. [Results and Discussion] Ablation studies and comparative experiments on the self-constructed dataset revealed significant performance gains. Ablation analysis: Reconstructing the backbone to LiteChimeraNet reduced FLOPs from 6.3 G (baseline) to 1.5 G, a decrease of 76.2%, while maintaining stable mean Average Precision (mAP@50), verifying its efficiency in removing background redundancy. Introducing WaveletPool significantly improved mAP@50 by 1.79 percentage points, confirming the critical role of explicitly preserving high-frequency components for serrated texture representation. Integrating SDL-Head further optimized feature alignment, increasing the recall rate to 94.47%. Overall performance: Flor-YOLO achieved a precision of 93.04%, recall of 94.47%, and mAP@50 of 96.10%. Compared to the YOLO11n baseline, these metrics improved by 3.52, 1.34, and 3.25 percentage points, respectively. Meanwhile, parameters and FLOPs were reduced by 51.2% to 1.26 M and 1.1 G (82.54% reduction). Flor-YOLO exhibited distinct advantages over YOLOv5n, YOLOv8n, YOLOv9t, YOLOv10n, and YOLOv12n in accuracy, mAP, and inference speed. Mechanism analysis: Spectral energy statistics showed that high-frequency energy intensified with increasing openness grades, aligning with the visual characteristics of petal expansion and wrinkle formation, thus validating the discriminative value of high-frequency information. Grad-CAM++ visualizations further validated that the improved model stably focused on petal edges and flower centers, demonstrating superior robustness over the baseline in complex backgrounds. [Conclusions] By constructing the LiteChimeraNet backbone, incorporating frequency-domain downsampling, and designing a detail-enhanced head, the proposed model effectively enhances the representation of critical details such as petal edges and flower centers while maintaining extremely low computational costs. Comprehensively, Flor-YOLO achieves an optimal balance between accuracy, model size, and real-time performance, demonstrating strong potential for deployment on low-power mobile terminals and embedded sorting equipment. Furthermore, the proposed frequency-aware lightweight design paradigm provides a valuable reference for other agricultural vision tasks relying on subtle textural differences.

Key words: Dianthus caryophyllus L. grading, object detection, Flor-YOLO, lightweight, wavelet pooling, frequency domain analysis