Reliable maturity detection of Sichuan pepper is critical for intelligent harvesting, precision agriculture, and automated quality assessment. However, existing vision-based methods often suffer from performance degradation in natural orchard conditions due to illumination variation, fruit occlusion, background interference, and substantial appearance diversity. In addition, the lack of dedicated Sichuan pepper maturity datasets limits the development and reliable evaluation of deep learning-based detection models. To address these challenges, this study proposes C-MorphYOLO (Circular-Morphological YO-LO), a morphology-aware detection framework for Sichuan pepper maturity recognition under field conditions. By incorporating circular morphological information into feature representation learning, the proposed framework enhances the perception of fruit structural characteristics and improves fine-grained maturity discrimination. Specifically, a Circular Depthwise Convolution Module (CDCM) is developed to strengthen boundary and shape-aware feature extraction, while a Feature Enhancement-based Upsampling Module (FEUM) and a Multi-Scale Adaptive Spatial Attention Gate (MASAG) are introduced to improve small-object representation and detection robustness. Furthermore, a natural-scene Sichuan pepper maturity dataset containing 2763 images and 2962 annotated instances was established for model training and comprehensive evaluation. Experimental results on the held-out validation set demonstrate that C-MorphYOLO achieves a Precision of 96.1% and an $\mathrm{m}\mathrm{A}{\mathrm{P}}_{50:95}$ of 89.1%, representing absolute improvements of 2.2% and 2.7%, respectively, compared with the YOLOv12s baseline. These results demonstrate that the proposed framework effectively improves maturity recognition and localization performance, providing a practical vision-based solution for intelligent Sichuan pepper harvesting and automated agricultural management.

Fine-grained maturity assessment of Sichuan pepper (Zanthoxylum bungeanum cv. Hanyuan) is challenging: CNNs are data-hungry and opaque, and vision-language models (VLMs) achieve only chance-level accuracy ($16.67\%$) on this six-way task when used end-to-end. We present KG-RAG, a retrieval-augmented classification framework that combines CNN visual features with a 25,881-triplet knowledge graph built from VLM-extracted structured attributes. KG-RAG introduces three key components: (i) a hierarchical attribute consistency score (HACS) that generalizes Jaccard re-ranking via mutual-information weighting and family-level regularization; (ii) a confidence-guided retrieval gate for adaptive parametric/non-parametric fusion, with calibration as a secondary benefit; and (iii) a two-stage VLM curriculum that repurposes a VLM—inaccurate as an end-to-end classifier but reliable as an attribute extractor—into a structured knowledge provider. On 2114 expert-annotated images (inter-annotator Cohen’s $\kappa =0.89$), $10\times 5$-fold cross-validation shows consistent gains across four backbones (+1.25 to +4.65 percentage points), with the largest gain in the low-data regime (+14.89 pp at 10% training data). Statistical significance is assessed via paired t-tests with repeated-CV caveats.