基于生成对抗网络和混合注意力机制残差网络的苹果病害识别

Apple disease recognition based on Wasserstein generative adversarial networks and hybrid attention mechanism residual network

准确识别并防治病害对提高苹果产量和质量具有重要意义。针对苹果病害图像因病斑区域小且易受背景干扰而导致识别准确率低的问题,设计一种基于混合注意力机制的残差网络(Convolutional Block Attention Module-Residual Network,CBAM-ResNet)模型。该模型在残差网络中嵌入串联的通道注意力模块和空间注意力模块,使模型将注意力集中于图像特征的病害区域,提高识别准确率。针对模型训练数据集样本数量不足且数据不均衡问题,提出基于Wasserstein距离的生成对抗网络(Wasserstein Generative Adversarial Networks,WGAN)对数据集进行扩充的方法。通过生成器与判别器的对抗训练生成10000张苹果病害图像,对CBAM-ResNet进行训练和测试,提高模型的泛化能力和鲁棒性。试验结果表明:与CNN、VGG-16、ResNet-50、Inception-V3等传统模型相比,CBAM-ResNet对苹果病害的识别效果更优,其识别准确率、精确率、召回率和F1-Score参数分别达到95.50%、95.40%、95.40%和0.95。该方法能够为苹果病害图像准确识别和实时监测提供技术支撑。

It is important to identify and control the disease accurately to improve the yield and quality of apples.Aiming at the problem of low recognition accuracy of apple disease images caused by small lesion areas and easy background interference,a Convolutional Block Attention Module-Residual Network(CBAM-ResNet)model was designed.This model embedded a tandem channel attention module and spatial attention module in the residual network so that the model can focus on the diseased areas of image features and improve the recognition accuracy.Aiming at the problem that the number of samples in the model training dataset is insufficient and imbalanced,this paper proposed a method for dataset expansion based on Wasserstein Generative Adversarial Networks(WGAN).Through the antagonistic training between generator and discriminator,10000 apple disease images were generated,and CBAM-RESNET was trained and tested to improve the generalization ability and robustness of the model.The experimental results showed that compared with traditional models such as CNN,VGG-16,ResNet-50,and Inception-V3,CBAM-ResNet had a better recognition effect on apple diseases,and its recognition accuracy,precision,recall,and F1-Score reached 95.50%,95.40%,95.40%,and 0.95.The method can provide technical support for accurate identification and real-time monitoring of apple disease images.