Abstract: Accurate detection of apple leaf diseases is of great significance for improving apple production and quality. Aiming at the problem that existing apple leaf disease detection models cannot make full use of information for given images, resulting in poor detection performance, an apple disease leaf detection based on multiscale feature fusion network was proposed. Vgg-16 network was firstly improved using depthseparable convolution, and improved network was used as a global feature extractor for apple leaf disease pictures. Secondly, Swin Transformer network was used as a local feature extractor. Next, the multiscale feature fusion network was proposed to fuse local and global features to construct multiscale features. Finally, the fusion multiscale features were used as input to a fully connected network for the detection of apple disease leaves. The experimental results showed that the proposed model could achieve 93.98% accuracy, 94.11% precision, 93.93% recall and 94.62% F1 value. Compared with the current mainstream apple disease leaf detection models, it was highly competitive in terms of detection accuracy and the amount of model parameters to be calculated.

Key words: apple disease, leaf detection, global feature, local feature, multiscale feature fusion network, disease identification

摘要： 准确地检测出苹果叶片的病害对提高苹果产量和质量具有重要意义。针对现有苹果叶片病害检测模型信息利用不充分，导致检测性能不佳的问题，提出一种基于多尺度特征融合网络的苹果病害叶片检测方法。首先，利用深度可分离卷积改进传统Vgg-16网络，并作为苹果叶片病害图片的全局特征提取器；其次，利用Swin Transformer网络作为局部特征提取器；然后，提出一种多尺度特征融合网络将局部和全局特征进行融合，构造多尺度特征；最后，将融合的多尺度特征作为全连接网络的输入，实现苹果病害叶片的检测。实验结果表明，所提出方法可以实现93.98%的准确率、94.11%的精准率、93.93%的召回率和94.62%的F1值。相比当前主流的苹果病害叶片检测模型，在检测精度和模型参数计算量等方面，均具有很强的竞争力。

关键词: 苹果病害, 叶片检测, 全局特征, 局部特征, 多尺度特征融合网络, 病害识别