BPSOF-YOLO11n: YOLO11n-based method for blueberry maturity detection using boundary protection and small-object features

Introduction Blueberries are small and densely clustered, and their color transitions between different stages of ripeness are relatively gradual, which may lead to detection omissions and confusion with adjacent classes in real-world scenarios. Methods We developed the BPSOF-YOLO11n model by adding a high-resolution P2 (the second layer of the feature pyramid, corresponding to a high-resolution feature map with an output stride of 4) detection branch to YOLO11n, inserting a Residual Convolutional Block Attention Module (ResCBAM) with a flexible gating mechanism into the P2 fusion path, and applying a Maturity Boundary Protection (MBP) classification loss during the second-stage fine-tuning process with a low learning rate. MBP combines hard semi-ripe positive sample learning, adjacent-class ranking, and endpoint class protection. We evaluated the model on public and in-house datasets using validation set and held-out test set metrics, matched second-stage control groups, error analysis, and LayerCAM visualizations. Results BPSOF-YOLO11n achieved mean average precision (mAP)@0.5 values ​​of 91.92% and 92.72% on the public and self-built validation sets, respectively. On the publicly available test set, compared to the regular second-stage fine-tuning, the three sub-comparisons of the matching improved by an average of 0.126 percentage points on mAP@0.5 and by an average of 0.099 percentage points on mAP@0.5:0.95. In the self-built seed-0 comparison, MBP improved the semi-mature AP@0.5 by 0.240 percentage points and reduced the adjacent-stage error from 38 to 35. Discussion BPSOF-YOLO11n enhances feature retention and ripeness classification capabilities for small fruits, while MBP provides task-specific optimization without increasing inference parameters or computational branches.