Abstract The characteristics of the complex background in crop disease image, the small disease area, and the small contrast between the disease region and the background that easily causes confusion between them, seriously affect the recognition robustness and accuracy. To address these issues, we propose a Self-Attention Convolutional Neural Network (SACNN), which extracts effective features of crop disease spots to identify crop diseases. Our SACNN includes a basic network and a self-attention network: the basic network is for extracting the global features of the image, and the self-attention network is for obtaining the local features of the lesion area. Extensive experimental results show that the recognition accuracy of SACNN on AES-CD9214 and MK-D2 is 95.33% and 98.0%, respectively. The recognition accuracy of SACNN on MK-D2 has outperformed the state-of-the-art method by 2.9%, which implies that the CNN with self-attention can focus on the important areas of the image, and thus can improve the recognition accuracy. Adding different levels of noise to the AES-CD9214 test set shows the anti-interference ability and the strong robustness of SACNN. In addition, we discuss the influence of the location selection, channel size setting, network number and other aspects of the self-attention network on the recognition performance, in order to show the self-attention network working mechanism and provide inspiration for future research.

中文翻译：

---

基于Self-Attention卷积神经网络的作物叶片病害识别

摘要 作物病害图像背景复杂、病区小、病区与背景对比度小易混淆等特点，严重影响了识别的鲁棒性和准确性。为了解决这些问题，我们提出了一种自注意力卷积神经网络（SACNN），它提取作物病斑的有效特征来识别作物病害。我们的SACNN包括基础网络和自注意力网络：基础网络用于提取图像的全局特征，自注意力网络用于获取病变区域的局部特征。大量实验结果表明，SACNN 在 AES-CD9214 和 MK-D2 上的识别准确率分别为 95.33% 和 98.0%。SACNN 在 MK-D2 上的识别准确率超过了 state-of-the-art 方法 2.9%，这意味着具有 self-attention 的 CNN 可以专注于图像的重要区域，从而可以提高识别准确率. 在 AES-CD9214 测试集上加入不同级别的噪声显示了 SACNN 的抗干扰能力和强大的鲁棒性。此外，我们讨论了自注意力网络的位置选择、通道大小设置、网络数量等方面对识别性能的影响，以展示自注意力网络的工作机制，为未来的研究提供启发。在 AES-CD9214 测试集上加入不同级别的噪声显示了 SACNN 的抗干扰能力和强大的鲁棒性。此外，我们讨论了自注意力网络的位置选择、通道大小设置、网络数量等方面对识别性能的影响，以展示自注意力网络的工作机制，为未来的研究提供启发。在 AES-CD9214 测试集上加入不同级别的噪声显示了 SACNN 的抗干扰能力和强大的鲁棒性。此外，我们讨论了自注意力网络的位置选择、通道大小设置、网络数量等方面对识别性能的影响，以展示自注意力网络的工作机制，为未来的研究提供启发。