BSD-YOLOv11: an improved YOLOv11n model for detecting brown-spot-like leaf lesions in Rosaceae fruit trees

Abstract

Brown-spot-like leaf lesions on Rosaceae fruit trees are typically small, irregular and indistinct. They are also easily confused with leaf veins, shadows and background textures, making them difficult to locate. This paper proposes BSDYOLOv11, a YOLOv11n-based model for detecting these lesions. The dataset consists of leaf images from publicly available apple, pear, and cherry trees.The YOLOv11n detection head is unchanged, alter three components in the backbone and neck. In the backbone, BTRD is used to replace some stride-2 downsampling convolutions to enhance the preservation of boundary transition information for small lesions; ShiftwiseConv is locally embedded into the high-level C3k2 module to form C3k2-SWConv, thereby improving the contextual representation of irregular lesion shapes; and in the neck, DPCF is introduced to replace conventional splicing-based fusion to enhance detail preservation during multiscale feature fusion. Comparative experimental results show that, under the same experimental settings, BSD-YOLOv11 achieved Precision, Recall, mAP50, and mAP50–95 of 85.85%, 80.60%, 86.75%, and 74.33%, respectively; while the original YOLOv11n achieved 83.96%, 78.28%, 85.14%, and 72.42%, respectively. Compared to the baseline, BSD-YOLOv11 achieved improvements of 1.89, 2.32, 1.61, and 1.91 percentage points, respectively, across the four metrics.