交叉注意力机制引导的无监督域自适应图像分类模型构建及其在细粒度实蝇识别中的应用

打开文本图片集

关键词：实蝇；无监督域自适应；注意力机制；图像分类

中图分类号：Q969.456.8

文献标识码：A

文章编号：1000-4440（2026）04-0756-07

Construction of an unsupervised domain adaptive image classification model guided by cross-attention mechanism and its application in fine-grained fruit fly recognition

PENG Yingqiong 1，2 ， RAO Yuxiang 1，2 ， LIAO Muxin 3 ， ZHONG Wenbo 1

School of Software， Jiangxi Agricultural University， Nanchang 330045， China; Key Laboratory of Agricultural Information Technology of Colleges and Universities in Jiangxi Province， Nanchang 330000， China; School of Computer Science and Engineering， Jiangxi Agricultural University， Nanchang 330045， China）

Abstract： Existing pest image classification models often suffer from performance degradation in cross-domain recognition. To address this issue， this study proposed an unsupervised domain adaptive image classification model based on focal regions， named FC-DroNet. First， the model introduced mask processing in feature extraction and utilized a cascaded cross-attention module to fuse horizontal， vertical， and global spatial features， thereby enhancing the ability to capture fine-grained local features. Meanwhile， a consistency constraint mechanism was introduced to suppress overfitting on the sourcedomain， thus improving cross-domain generalization performance. A dataset， FD4Set， containing four types of images， namely Bactrocera cucurbitae， Bactrocera scutellata， Bactrocera tau， and Bactrocera dorsalis （Hendel）， was constructed to verify the performance of the FC-DroNet model. The results showed that the precision of the FC-DroNet model on the test set reached 99.41% and the F1 reached 99.22%， both higher than those of the ResNet50 model， AlexNet model， VGG-16 model， LeNet-5 model， ConvNext model， and MobileVit model. The findings of this study provide technical support for the intelligent identification of field pests.

Key words： Bactrocera; unsupervised domain adaptation; attention mechanism; image classification

图像分类是计算机视觉的基础任务，旨在通过分析图像内容将其准确归类。（剩余10987字）