基于注意力机制非对称残差网络和迁移学习的玉米病害图像识别

为实现玉米病害图像快速、准确识别,提出了一种基于非对称注意力机制残差网络(asymmetric convolution attention resnet,ACA-Resnet)的图像检测模型.在残差网络的基础上,通过引入非对称卷积结构减少参数量,加快模型训练速度,同时引入注意力机制,改善模型的表达能力,提高检测准确率.为减小由于病害图片数量不足而造成的过拟合现象,采用迁移学习的方法提高模型的稳定性和泛化能力.结果 表明,ACA-Resnet经过ImageNet数据集预训练后对玉米病害图像的平均识别准确率可达到97.25％,较VGG-16、Inception-V3和ResNet50等网络分类效果更好,相较于Resnet50训练速度明显提升.可见本文方法训练速度快,识别精度高,可为玉米病害检测提供借鉴.

Image Recognition of Maize Disease based on Asymmetric Convolutional Attention Residual Network and Transfer Learning

For realizing the fast and accurate identification of corn disease images, an image detection model based on the Asymmetric Convolutional Attention Resnet (ACA-Resnet) is proposed. Based on the residual network, the asymmetric convolutional structure is introduced to reduce the number of parameters and accelerate the training speed of the model. Meanwhile, the attention mechanism is employed to improve the expressing ability of the model and the detection accuracy. In order to circumvent the overfitting problem caused by the insufficient disease images, the transfer learning algorithm is adopted to enhance the stability and generalization ability of the model. The average recognition accuracy of ACA-Resnet after pre-training with ImageNet data set reached 97.25%, which was better than that of VGG-16, Inception-V3, and Resnet50, and the training speed was significantly improved compared with Resnet50. The method presented in this paper has fast training speed and high identification accuracy, which can provide reference for the detection of maize disease.