自然场景下获取的黄瓜病害图像存在光照、 土壤等噪声, 严重影响黄瓜病害识别精度, 现有检测模型占用内存较大, 难以实现黄瓜病害的实时检测。 以自然环境中黄瓜霜霉病、 白粉病和炭疽病3种病害的可见光光谱图像为研究对象, 提出一种基于可见光谱和改进YOLOv5目标检测网络的黄瓜病害识别模型, 探索自然环境中黄瓜病害的准确实时检测并降低检测模型存储成本的方法。 为平衡检测精度和模型所需存储空间, 以轻量级网络结构YOLOv5s为基线模型, 引入SE注意力机制, 提取特征维度信息, 降低复杂背景对检测结果的影响, 提高模型的检测精度; 引入深度可分离卷积替换基线模型中的标准卷积, 减少模型参数带来的计算负担, 提高检测速度。 检测模型接收任意像素的可见光光谱图像并调整成640×640像素作为检测网络输入, 输出表示黄瓜病害发生区域及病害类别, 使用COCO数据集上预训练权重初始化网络权重。 实验结果表明, 改进后的YOLOv5s-SE-DW模型对黄瓜霜霉病、 白粉病和炭疽病的检测精度分别达到了78.0%, 80.9%和83.6%, mAP高达80.9%, 模型存储空间仅为9.45 MB, 浮点运算次数仅为11.8 G, 相比基线模型mAP提高了2.4%, 运算次数减少了4.6 G, 模型所需的存储空间降低了4.95 MB, 在减小模型所需内存的同时提升病害检测准确率; 进一步与经典两阶段目标检测网络Faster-RCNN和单阶段目标检测网络YOLOv3, YOLOv3-tiny, YOLOv3-SPP以及YOLOv4进行对比, 提出的YOLOv5s-SE-DW模型相比对比模型中表现最优的YOLOv4模型mAP提高了3.8%, 检测时间和存储空间大幅度降低。 综合结果表明, 所提出的YOLOv5s-SE-DW网络对于自然场景中的黄瓜病害检测具有良好的精度和实时性, 能够满足黄瓜实际种植环境中病害检测的需求, 为实际应用场景下黄瓜病害自动检测的实现提供参考。

The cucumber disease images acquired in natural scenes have noise, such as light and soil, which seriously affects the accuracy of cucumber disease recognition. The existing detection models occupy a large memory, making it difficult to achieve real-time detection of cucumber diseases.The visible spectral images of three diseases of cucumber, namely downy mildew, powdery mildew and anthracnose, in the natural environment are used as the research object. In this paper, a cucumber disease identification model based on the visible spectrum and an improved YOLOv5 object detection network is proposed to explore the accurate real-time detection of cucumber diseases in the natural environment and to reduce the storage cost of the detection model. The lightweight network structure YOLOv5s is used as the baseline model. The SE attention mechanism is introduced to extract the feature dimensional information to reduce the influence of complex background on the detection results and improve the detection accuracy of the model. The depth separable convolution is introduced to replace the standard convolution in the baseline model to reduce the computational burden caused by the model parameters and improve the detection speed. The network receives visible spectral images of arbitrary pixels and adjusts them to 640×640 pixels as input, outputs the cucumber disease occurrence region and disease category, initializes the detection method and trains the detection network using pre-trained weights on the COCO dataset.The experimental results show that the improved YOLOv5s-SE-DW model achieves 78.0%, 80.9%, and 83.6% detection accuracy for cucumber downy mildew, powdery mildew, and anthracnose, respectively, with mAP as high as 80.9%. The storage space of the model is only 9.45 MB, and the number of floating point operations is only 11.8 G. Compared with the baseline model, the mAP is improved by 2.4%, 4.6 G reduces the number of floating point operations, and the storage space required for the model is reduced by 4.95 MB. The improved model improves disease detection accuracy while reducing the storage memory. Further comparison with the classical two-stage target detection network Faster-RCNN and single-stage target detection networks YOLOv3, YOLOv3-tiny, YOLOv3-SPP, and YOLOv4 shows that the proposed YOLOv5s-SE-DW model improves the mAP by 3.8% compared with the best-performing YOLOv4 model among the comparison models, and the detection time and storage space are significantly reduced. The detection time and storage space are substantially reduced. The comprehensive results show that the proposed YOLOv5s-SE-DW network has good accuracy and real-time performance for cucumber disease detection in natural scenarios, which can meet the demand for disease detection in actual cucumber growing environments and provide a reference for cucumber disease detection in practical application scenarios.