YOLO12-SDS: an enhanced lightweight model for accurate and real-time fig maturity detection.

The contents of active ingredients such as isoprenyl phenols and the quality of figs are significantly regulated by ripeness. Accurate ripeness grading is critical for quality improvement and efficiency enhancement in the fig industry. However, existing detection methods suffer from poor environmental adaptability and excessive model redundancy, failing to simultaneously satisfy the requirements of high-precision, lightweight and real-time detection. To address the above limitations, this study takes YOLO12n as the baseline model and proposes a lightweight detection model named YOLO12-SDS for fig ripeness grading. Structurally, three customized modules are integrated into the proposed model. The SPDSPPF module adopts spatial rearrangement and multi-scale pooling fusion to strengthen the retention of small-object details and high-level feature extraction capability. The DySample module employs the PL-based learnable upsampling strategy to restore fruit contour, color and texture information with superior performance. The SCSA module leverages the collaborative spatial-channel attention mechanism to highlight critical ripeness-related features and suppress interference from complex backgrounds. Experimental results on the fig dataset demonstrate that compared with the original YOLO12n model, YOLO12-SDS achieves performance improvements of 2.1, 6.6, 4.8 and 1.0 percentage points in Precision, Recall, mAP@50 and mAP@50:95, respectively. The proposed model has a parameter count of 2.74 M, a computational cost of 6.0 GFLOPs and an inference speed of 115.6 FPS. Embedded deployment and testing are completed on the CUBE NANO platform. This study provides a technical solution with high precision, strong real-time performance and favorable deployment adaptability for fig ripeness grading in greenhouse facilities, which possesses great practical application value and broad promotion prospects.