Tiny target detection algorithm in field wheat canopies based on MBCD-YOLO

To address the detection challenges posed by tiny wheat-ear targets, highly similar wheat canopy backgrounds, dense occlusion, and scale variation in low-altitude UAV images, this paper proposes MBCD-YOLO, a lightweight detection algorithm based on YOLO11n. First, a Micro-Feature Preservation Module (MFPM) uses Space-to-Depth rearrangement to preserve fine-grained spatial information during downsampling. Second, a Boundary-Semantic Synergy Block (BSSB) fuses fixed Sobel-gradient priors with learned semantic features to improve boundary discrimination under homogeneous backgrounds. Third, a Context-Aware Reconstruction Module (CARM) uses efficient up-convolution composed of 2x upsampling, 3 × 3 depthwise convolution, channel shuffle, and 1 × 1 pointwise fusion to reduce feature aliasing and spatial misalignment in multi-scale fusion. Finally, a Distribution-Guided Quality Head (DGQH) introduces localization-quality estimation derived from regression distributions to improve dense-object ranking during NMS. Experiments on the Global Wheat Head Detection Dataset (GWHD) show that MBCD-YOLO achieves 50.6% mAP@50–95 and 81.7% recall while maintaining 89 FPS. In the counting-oriented evaluation, MBCD-YOLO achieves an MAE of 4.92, an RMSE of 7.18, and a counting accuracy of 88.33%, demonstrating its effectiveness for dense wheat-ear detection and counting.