Abstract: Aiming at the problems of low accuracy of tomato disease identification due to overlapping leaves and small targets in complex environments, a multiscale cascade model (IMS-Cascade) is proposed. The model is based on cascade neural network (Cascade R-CNN), the switchable Atrous convolution of fused context information is introduced into the backbone network, and complex multiscale convolution kernels are used to extract target features to solve the problem that the shape of the same disease is greatly different due to leaf occlusion, and the feedback connection module is added to the feature pyramid networks, so that the model can extract features for many times and improve the utilization of shallow information. Finally, the gradient of accurate samples is increased in the loss function to reduce the influence of abnormal samples on the model. When the model is applied to a portion of the tomato leaf disease dataset published by Plant Village, the mean average precision (mAP) reaches 89.1% and the average precision reaches 99.36%. These results represent improvements of 2.5% and 1.84%, respectively, over the original Cascade R-CNN model. This indicates higher detection accuracy, which is beneficial for tomato disease detection in complex environments.

Key words: tomato disease detection, feedback connection, feature pyramid networks, dilated convolution, multiscale

摘要： 针对复杂环境下因叶片重叠遮挡以及目标较小等原因而导致番茄病害识别准确率较低的问题，提出一种多尺度级联模型(IMS-Cascade)。该模型以级联神经网络(Cascade R-CNN)为基础，在主干网络中引入融合上下文信息的可切换空洞卷积，使用复杂的多尺度卷积核提取目标特征，解决同种病害因叶片遮挡而形状差异较大的问题，并在特征融合网络中添加反馈连接模块，使模型可以进行多次的特征提取，提高浅层信息的利用率。最后在损失函数上增大准确样本的梯度，降低异常样本对模型的影响。将该模型用于Plant Village公开发表的部分番茄叶病害数据集上，mAP达到89.1%，平均准确率达到99.36%，分别比原始Cascade R-CNN模型提高2.5%和1.84%，具有更高检测精度，有利于复杂环境下的番茄病害检测。

关键词: 番茄病害检测, 反馈连接, 特征金字塔网络, 空洞卷积, 多尺度