为解决田间环境下由于叶片间遮盖和堆叠等因素引起的木薯叶病害识别困难的问题,本文提出一种基于改进YOLOX网络的木薯叶病害检测(Cassava leaf disease detection,CDD)模型.首先,对复杂背景下木薯叶病害图像数据集进行数据增强,以减少环境影响造成的识别困难.其次,在YOLOX网络的基础上,使用多尺度特征提取模块加强细粒度特征提取并降低模型计算量,同时嵌入通道注意力机制,提高网络的表征能力.最后,结合质量焦点损失函数作为分类损失函数辅助网络收敛,提高目标分类的准确性.实验结果表明,提出的CDD模型对复杂背景下木薯叶病害进行检测,网络参数量为5.04 ×106,平均精度均值达93.53%,比基础模型高6.02个百分点,综合检测能力优于多种主流模型.因此,本文提出的CDD模型对田间木薯叶病害具有更快更准确的检测能力,为实现农作物病害检测提供了可借鉴的方法.
The present method has some difficulties in recognizing cassava leaf diseases in a field environment,such as covering and stacking between leaves.Based on the YOLOX network,cassava leaf disease detection(CDD)model was proposed.Firstly,the cassava leaf disease image data under complex background was augmented to reduce the recognition difficulty caused by environmental impact.Secondly,built on the YOLOX network,the lightweight multi-scale feature extraction(LME)module was used to strengthen fine-grained feature extraction and reduce the amount of model calculation.At the same time,the channel attention mechanism was embedded to improve the representation ability of the network.Finally,the quality focal loss was used as a part of the classification loss to assist the network convergence and improve the accuracy of target classification.In conclusion,the proposed CDD model can detect cassava leaf disease under complex background.The amount of network parameters was 5.04 x 106 and the mean average precision was 93.53%,which was 6.02 percentage points higher than that of the non-optimized network model.Comprehensive detection ability was better than that of previous models.Therefore,the proposed method CDD had faster and more accurate detection ability for cassava leaf diseases in the field,and provided a reference method for realizing intelligent field detection.