Authors

Abstract

Accurate detection of grape diseases in controlled facility environments is of significant importance. In this study, a novel detection model, PWE-YOLO, is proposed based on YOLO v11. In this model, the C3k2 modules in the Neck network are replaced with C3k2-PConv modules, which maintain high-precision feature extraction while reducing model parameters and memory consumption. Furthermore, a WaveletPool module is introduced to replace all Conv modules in the model, except for the 0th and 1st layers, enhancing feature representation and detection accuracy for grape diseases while further reducing model parameters and floating-point operations. EMA attention modules are incorporated at the 17th and 24th layers to improve small-object detection capabilities, thereby increasing overall detection precision. Experimental results demonstrate that PWE-YOLO achieves a precision of 85.0%, recall of 83.6%, and mAP of 86.9%, with 1.96 × 10⁶ parameters and 5.1 × 10⁹ floating-point operations. Compared with the original YOLO v11 model, precision, recall, and mAP increase by 4.5, 1.4, and 1.9 percentage points, respectively, while parameters and floating-point operations decrease by 24.03% and 10.05%. Relative to YOLO v8, v9, v10, v12, and v13, precision improves by 2.2%, 5.3%, 3.3%, 2.2%, and 2.5%, recall by 6.0%, 2.8%, 5.2%, 8.8% and 3.8%, and mAP by 1.9%, 1.8%, 2.8%, 5.0%, and 2.3%, respectively, with corresponding reductions in model parameters and floating-point operations. These results indicate that PWE-YOLO not only provides high-accuracy detection for grape diseases but also reduces computational complexity, achieving a lightweight architecture and faster detection speed, making it suitable for deployment in resource-constrained target detection scenarios.

Abstract in Chinese

为实现设施环境中葡萄病害目标的准确检测，本文在YOLO v11的基础上，提出一种适用于设施环境葡萄病害目标检测模型PWE-YOLO。将颈部网络中的C3k2模块替换为C3k2-PConv模块，在保持高精度特征提取能力的基础上，降低模型参数量和内存占用量；同时在模型中引入WaveletPool模块，替换模型中除第0层、第1层外的所有Conv模块，提高葡萄病害的特征表达能力与检测精度，进一步降低模型参数量和浮点运算量；且在模型的第17层和第24层中添加EMA注意力机制模块，增强模型对小目标的检测能力，提升模型的检测精度。结果显示，PWE-YOLO模型的精确率为85.0％、召回率为83.6％、mAP为86.9％、参数量为1.96 × 10⁶、浮点运算量为5.1 × 10⁹。与原始YOLO v11模型相比，该模型的精确率、召回率和mAP分别提高4.5、1.4、1.9个百分点，参数量和浮点运算量分别降低24.03%、10.05%。与YOLO v8、YOLO v9、YOLO v10、YOLO v12、和YOLO v13模型相比，其精确率分别提高2.2、5.3、3.3、2.2、和2.5个百分点，召回率分别提高6.0、2.8、5.2、8.8、和3.8个百分点，mAP分别提高1.9、1.8、2.8、5.0、和2.3个百分比，参数量分别减少34.88%、25.19%、27.40%、21.91%和20.00%，浮点运算量分别减少37.03%、52.34%、37.80%、12.07%和17.74%。结果表明，本文提出的PWE-YOLO模型在葡萄病害检测任务中不仅能够提供高准确性的检测性能，同时也有效降低了计算复杂度，具有较轻的模型结构和较快的检测速度，适合应用于对计算资源要求较高的目标检测任务中。