Abstract: In response to the demand for real-time, high-performance discrimination of tobacco curing stages in intelligent bulk curing barns, based on the stage variation characteristics of leaf color and morphology during the curing process, this paper proposes a multi-module collaboratively optimized model based on YOLOv8n, named RC-SL-WI-YOLOv8n, which can improve the recognition accuracy and real-time performance of the model for the tobacco curing stage and meet the requirements of intelligent management for precise curing. First, the standard convolution Conv modules in the backbone network are replaced with RFCBAMConv to achieve spatially adaptive receptive field adjustment, enabling the model to flexibly focus on feature differences across various regions of tobacco leaves. Second, a ColorAttention module is designed and introduced to enhance the extraction of key color features. Third, the SPPF module is extended to SPPF＿LSKA, which enlarges the global receptive field and strengthens long-range dependency modeling, thereby improving the model's ability to perceive overall color and morphological trends of tobacco leaves. Finally, a combined loss function integrating Wise-IoU and Inner-IoU is adopted to further improve the recognition accuracy and robustness for large targets. Compared with the original YOLOv8n model, the proposed RC-SL-WI-YOLOv8n achieves a precision of 92.52%, a recall of 91.51%, and an F₁-score of 91.77% on the test set, corresponding to improvements of 10.11, 10.25, and 10.48 percentage points, respectively, demonstrating significant performance enhancement. Moreover, in the test environment, the RC-SL-WI-YOLOv8n model has a parameter count of 3.09 M, a computational load of 3.95 GFLOPs, and an inference speed of 93 FPS, which can meet the application requirements of lightweight deployment. This study validates the potential of the YOLOv8n model for tobacco leaf state discrimination in complex curing scenarios, and achieves substantial performance gains through multi-module collaborative optimization, providing a key reference and feasible path for developing high-performance real-time discrimination lightweight models for this task.