Addressing the challenges of complex background interference, low lighting conditions, small target recognition, and difficulty in maturity grading in the automated detection of Pleurotus ostreatus, this study proposes a lightweight improved scheme based on color feature enhancement. By collecting 4779 images from five developmental stages in three typical planting environments, including greenhouses and mushroom houses, an HSV hue analysis database was established to determine key hue intervals [4°, 38°] or [110°, 155°] for different environments. Secondly, based on the hue interval distribution of Pleu-rotus ostreatus, YOLOv13 was used as the base model, with the addition of an HSV hue mask as the fourth channel to improve the input layer. The custom ColorWeight module was used to enhance color feature expression; the hypergraph computation module was improved to enhance feature correlation; and the neck network incorporated the StockenAttention module to improve the ability to capture maturity features. The accuracy of the improved model was increased to 89.5% in mAP@0.5 (+3.3%), surpassing the mainstream YOLOv8n-12n series. Efficiency optimization achieved real-time detection at 12.58 FPS on the RTX3090Ti platform. In practical applications, the accuracy of maturity recognition was significantly improved, with a 73.6% decrease in the misclassification rate of maturity and a reduction in missed detections, achieving an F1 score of 0.91. In conclusion, through the deep integration of Hue features and deep learning models, while ensuring lightweight deployment (with only a 10.5% increase in parameter count), the accuracy and practicality of Pleurotus ostreatus detection were significantly improved, providing an effective solution for intelligent mushroom house management.

Harvest-window recognition of facility-grown Pleurotus ostreatus is challenged by clustered fruiting bodies, occlusion, and subtle differences between near-mature and mature stages. We developed a two-stage framework combining YOLO26s detection with a lightweight dual-view classifier, RepViT-RSG-BEHarvestNet. YOLO26s was selected from four candidate detectors using the validation subset of a 2946-image dataset. Matched detections produced 3547 paired target and context regions of interest. The classifier combined stage-guided expert routing with boundary soft labels and teacher distillation. On the independent test set, YOLO26s achieved a Precision of 0.9681, Recall of 0.9117, and mAP50 of 0.9787. Across ten paired runs on matched test regions, the final classifier achieved a Macro F1 of 0.8827 ± 0.0044, compared with 0.8589 ± 0.0054 for the dual-view baseline, while adding 0.071 M parameters. On the near-mature–mature subset, Macro F1 increased from 0.7994 ± 0.0115 to 0.8563 ± 0.0105. On all 296 test images containing 404 annotated targets, the complete pipeline achieved target-level Precision of 0.7770, Recall of 0.7941, a false harvest rate of 0.1007, and a missed harvest rate of 0.2059. Mean end-to-end latency was 94.7 ms per image. The results show improved boundary classification while identifying detection errors as a remaining constraint on complete-image harvest decisions.

Accurate maturity recognition of Pleurotus citrinopileatus fruiting bodies is important for growth monitoring and harvest assistance in controlled cultivation. However, multiple developmental stages often coexist within one image, and near-mature fruiting bodies share visual characteristics with both adjacent stages. We developed a two-stage RGB-based framework for instance-level maturity recognition in complex cultivation scenes. After fruiting-body localization, each detected instance was represented by paired local and context views and classified using a DINOv3 ViT-S/16-based ordinal model. The task-specific methodological contribution is a morphology-guided fusion module that incorporates automatically extracted geometric, color, intensity, and edge descriptors into the dual-view visual representation. Object-level matching was used to quantify how detection and classification errors affected the complete pipeline, and temperature scaling was applied for probability calibration. Among five candidate detectors, RF-DETR-Small achieved validation mAP50 and mAP50–95 values of 0.9747 and 0.8813, respectively. On the independent ROI test set, the final classifier achieved an accuracy of 0.8779, a macro F1 of 0.7950, a near-mature F1 of 0.5546, and a QWK of 0.9068. Temperature scaling reduced ECE from 0.1094 to 0.0564 without changing the predicted labels. On 370 original cultivation images containing 475 ground-truth instances, the unfiltered pipeline achieved an object-level macro F1 of 0.7072 and a near-mature F1 of 0.4672. Candidate-box filtering increased these scores to 0.7305 and 0.4885, respectively, but increased the number of missed instances from 14 to 23. These results demonstrate the feasibility of non-destructive instance-level maturity monitoring while identifying near-mature discrimination and candidate-box control as priorities for practical application.