基于改进U-Net网络的多尺度番茄病害分割算法

Multi-Scale Tomato Disease Segmentation Algorithm Based on Improved U-Net Network

针对当前农作物病害分割与识别模型病斑分割精度低、数据集不充分、训练速度过慢等问题,构建了一种基于改进的U-Net网络多尺度番茄叶部病害分割算法。在U-Net网络结构基础上进行改进,减小图像输入尺寸,在编码器中使用非对称Inception多通道卷积替换传统卷积,实现多尺度提取病害特征,提升模型准确度;在解码器中加入注意力模块,关注番茄病害边缘,减小上采样噪声;引入GN加速模型收敛,并将改进U-Net网络用在PlantVillage数据集上进行预训练,提高模型的分割准确度和速度。改进后的方法准确率、召回率和MIoU分别为92.9%、91.1%、93.6%,实验结果表明,该方法能够有效地提高模型对番茄的病害分割性能。

Aiming at the problems of current crop disease segmentation and recognition models with low disease spot segmentation accuracy, insufficient dataset and too slow training speed, a multi-scale tomato leaf disease segmentation algorithm based on improved U-Net network is constructed. This model makes improvements on the basis of the U-Net network structure, reduces the image input size, uses asymmetric Inception multi-channel convolution in the encoder to replace the traditional convolution, which can realize multi-scale extraction of disease features and improve the accuracy of the model. The decoder adds attention module to pay attention to the edge of the tomato disease and reduce the upsampling noise. Moreover, this model introduces GN to accelerate its convergence, and using the improved U-Net network to pretrain on the PlantVillage dataset can improve the segmentation accuracy and speed of the model. The accuracy, recall and MIoU of the proposed method are respectively 92.9%, 91.1% and 93.6%. The experimental results show that this method can effectively improve the model’s performance of tomato disease segmentation.