耦合全局与局部特征的苹果叶部病害识别模型

为充分利用苹果叶部病害图像类间差异小且类内差异大的特点，该研究基于全局与局部特征的交互式耦合对特征提取方法进行了优化，设计出一种苹果叶部病害识别模型。首先，在全局特征提取分支设计了一个注意力融合模块，以融合通道和空间上的信息而增强卷积提取到的特征图，并由增强后的特征图生成全局特征以及注意力激活图；然后，在局部特征提取分支，利用注意力激活图的引导，设计了一个裁剪模块对原图像进行裁剪，以得到可能包含病害信息的图像块且嵌入生成局部特征；最后，通过设计多头交叉注意力特征耦合模块，实现全局特征和局部特征的双向交叉耦合。基于苹果病害图像数据集的试验结果表明，将全局与局部特征进行交互耦合能有效提升模型对苹果叶部病害图像的特征提取能力，其识别准确率可达到98.23%，且较之单纯的局部或全局特征提取分支，准确率分别提高了3.39与4.61个百分点，所提模型可用于实现自然场景下的苹果叶部病害自动识别。

Apple leaf disease identification model by coupling global and patch features

Abstract: Apples in China accounts for more than 50% of the global production and consumption at present. However, the quantity and quality of apples have been threaten by the various diseases, such as alternaria boltch, brown spot, mosaic disease, gray spot, and rust. The CNN-based methods can be expected to recognize the crop leaf disease for the high recognition rates. But, there is still lacking on the recognition accuracy, due mainly to the lack of linkage between global and patch features of disease images in the general disease recognition models. In this study, a disease recognition model was proposed using the patch and global features interactively coupling model (GPF-IC). The main characteristics were also addressed for the small inter-class and large intra-class differences in the apple leaf disease images under natural conditions. Firstly, an attention fusion module was designed in the global feature extraction branch. The convolutionally extracted feature maps were then enhanced to fuse the information on the channels and spaces. The global features and attention activation maps were generated from the enhanced feature maps. Secondly, a cropping module was designed to crop the original image using the attention activation maps. The blocks of images were obtained with the disease information in the patch feature extraction branch, particularly with the patch features. Thirdly, the multi-head cross-attention feature coupling module was designed to realize the bi-directional cross-coupling of patch and global features. As such, the recognition accuracy was improved to enhance the representation capability of fine-grained features of disease images. Finally, three operations of data enhancement were used to evaluate the learning effect of the model for the less overfitting, due to the insufficient training data. A total of 30,540 disease images of six types of apple leaves were obtained with the sufficient number of samples and balanced distribution. The improved model was included as follows. 1) The global feature extraction branch was proposed to promote the disease recognition accuracy by 0.79 percentage points using the attention fusion module. 2) A patch feature extraction branch and a cropping module were introduced to implement the local feature extraction. The model accuracy was then improved by 1.22 percentage points than before. 3) A multi-head cross-attention coupling module was proposed to couple the features from the different branches for the feature extraction and expression capability of the model. The recognition accuracy was improved by 3.39 percentage points, which was the highest recognition accuracy of 98.23%. The experiment demonstrated that the feature extraction can effectively exclude the non-target noises to locate the most discriminative region using the global feature extraction branch. The patch feature extraction branch was efficiently acquired the patch information using the image block embedding. The feature coupling module was realized the interactive coupling of global and patch tokens for the better fine-grained feature representation using multi-headed cross-attention. The GPF-IC was achieved in the 98.23% recognition accuracy of apple leaf disease. The finding can provide a technical support for the automatic recognition of apple leaf diseases in natural scenes.