MULTI-SCALE CROSS-FUSION FINE-GRAINED NETWORK FOR IDENTIFYING INVASIVE PLANTS

Hang Sun, Yuting Zang, Lu Wang, Shun Ren, Xidong Wang, and Xiaolin Chen

Keywords

Fine-grained recognition, multi-scale cross-fusion, centre loss, invasive plant identification

Abstract

The invasion of alien plants has resulted in detrimental impacts on the ecological, economic, and societal aspects. The current fine-grained identification methods have not considered features of different scales, which is crucial for extracting more discriminative features. Additionally, the presence of small inter-class differences and large intra-class differences in fine-grained classification tasks significantly heightens the challenges of accomplishing accurate fine- grained classification. To address these issues, we have designed a dual-branch cross-fusion fine-grained network to integrate different- scale features and enhance the discriminative ability of deep learning features. Specifically, we have developed a multi-scale cross-fusion attention module to fuse features of different scales and retain the most important areas for classification and recognition. Moreover, we have employed a simple and effective centre loss on the dual-branch network to obtain deep features with key learning objectives, namely, inter-class distribution and intra-class compactness. Experimental results on the iNat2021-Plants, iNat2018-Plants, and FGVC-Aircraft datasets demonstrate that the proposed method achieves recognition accuracies of 76.8%, 73.8%, and 93.6%, respectively. This indicates that the method is capable of more precise fine-grained recognition and provides new insights for AI-assisted invasive plant identification systems.

Important Links:

Go Back