Abstract: Fusarium Head Blight (FHB) of wheat is one of the most terrible diseases that lead to wheat yield reduction. It is of great significance to carry out automatic identification research on wheat FHB. However, the traditional methods of segmentation and recognition of wheat scab under complex field background is generally carried out through threshold value, color histogram, etc., and its segmentation and recognition accuracy is poor and its generalization ability is not satisfactory. In order to quickly and accurately segment FHB scab and then effectively confirm the severity of the disease for assisting agricultural workers to carry out subsequent researches, this paper proposes a semantic segmentation network model, UNetA, based on UNet structure and attention mechanism for wheat FHB. The wheat ear pictures are augmented and then input into the convolution layer of UNetA model for extracting feature map. The Attention mechanism consists with position attention and channel attention after convolution layer to make further extraction, and then EncoderDecoder structure with skipconnection and BN layer makes up the rest part together. The whole network utilizes crossentropy loss with weighted parameters to balance the gap between classes and to measure the difference between predicted label and actual label. Subsequently, UNetA model is compared with the stateoftheart methods. The experiment results show that the proposed method performs favorably against others in terms of MIoU in the same configuration and obtains an 83.90% of MIoU. Moreover, the proposed method spends 0.588 0 s time for wheat sacb segmentation, shorter than others.

Key words: UNet, attention mechanism, fusarium head blight of wheat, image sematic segmentation

摘要： 小麦赤霉病是导致小麦大幅度减产的病害之一，对其开展自动识别研究具有重大意义。然而，传统方法一般通过阈值、色彩直方图等在农田复杂背景下开展小麦赤霉病的分割识别研究，其分割识别精确度较差并且泛化能力也不尽如意。为了在节省大量人力成本的同时对小麦赤霉病病斑进行快速、准确地分割从而辅助农业工作者对小麦患病的严重程度进行确认并开展后续的针对性研究，提出一种融合卷积神经网络和注意力机制的小麦赤霉病语义分割网络模型UNetA。该模型依据小麦赤霉病数据集的特点，使用融合了位置自注意力和空间注意力的注意力机制模块，并将注意力机制模块融入改进了的UNet结构中，再利用加权交叉熵损失函数来衡量预测值与实际值的差距同时缓解样本不均衡问题。试验结果表明，与现有的经典网络模型相比，UNetA模型的分割精度和实时性明显占据优势，其MIoU值达到83.90%，分割单张图像所用平均时间仅为0.588 0 s。

关键词: UNet, 注意力, 小麦赤霉病, 图像语义分割