YOLOv11-DAFNet: A multi-directional attention and multi-dilated residual fusion network for cucumber downy mildew sporangia detection in microscopic images

To address the challenges in microscopic image detection of cucumber downy mildew sporangia, where extremely small target sizes and pronounced geometric deformations hinder detection performance, this study proposes a novel detection model, YOLOv11-DAFNet. First, to overcome the limitations of conventional convolutions in accurately representing edge pixels and capturing multi-directional features of sporangia due to fixed convolutional orientations, a Directional Attention Fusion Convolution (DAFConv) is designed, which effectively enhances the modeling capability for multi-directional edge features. Second, in response to the limited receptive field of the original C3k2 module and its insufficient adaptability to the multi-scale deformation of sporangia, a Multi-Dilated Residual Block (C3k2_MDRBlock) is introduced. By embedding multiple dilation rate convolution kernels into the C3k2 structure, this module strengthens the network’s ability to perceive complex shapes and multi-scale features. Finally, to reinforce the interaction and fusion between shallow and deep features, a Context-Aware Attention Fusion Module (CAFMFusion) is proposed, which integrates a context-aware mechanism with pixel attention to adaptively generate fine-grained pixel-level weights, thereby enabling precise, dynamic spatial fusion and selection between shallow and deep features. Experimental results demonstrate that YOLOv11-DAFNet achieves an [email protected] of 88.2% on the test dataset, representing a 6.9 percentage point improvement over the baseline YOLOv11n, while precision and recall are increased to 90.7% and 85.4%, respectively. This work provides a fast and accurate detection approach for cucumber downy mildew sporangia.