A non-invasive learning branch to capture leaf-image attention for tree species classification

• Published in: Multimedia tools and applications Vol. 81; no. 10; pp. 13961 - 13978
• Main Authors: Song, Yupeng, He, Fazhi, Liu, Yanan
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.04.2022; Springer Nature B.V
• Subjects: Artificial neural networks; Classification; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Datasets; Discrimination; Image classification; Learning; Multimedia; Multimedia Information Systems; Similarity; Special Purpose and Application-Based Systems; Species classification; Trees; Tree species; Computer vision; Attention mechanism; Image classification
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-022-12036-6

Tree species classification is a necessary and challenging task for both multimedia technique and forestry engineering. Continuous developments in multimedia and vision technology enable this task to be achieved faster and more conveniently without relying on professional equipment. However, encountering many species and high similarity tree species, the existing methods cannot be completed well. In this paper, we propose a non-invasive attention learning branch structure to strengthen the discrimination of CNNs for complex tree species classification. Firstly, we proposed a novel network, namely Attention Learning Convolutional Neural Network (ALNet), in which a non-invasive learning branch for leaf-image attention capture is added in the typical convolutional blocks to strengthen the discrimination capacity of CNN, especially in the challenging task of classification of tree species with high similarity leaves, leaves image features only slightly different. Secondly, we designed a novel Attention Learning Branch (ALB), which can capture and learn leaf image attention, and be non-invasively integrated into existing CNN pipelines to capture the Region of Interest (ROI) of input features for image classification processing. Thirdly, we adopted a reliable residual approach to complete the integration of network branches, which improved overall network performance. In the Leafsnap dataset, the ALNet test was performed with 91.76 % accuracy on the classification of 185 species of trees, in which trees of different subspecies contained highly similar leaf images. Moreover, in the conventional Flavia dataset of tree leaf-image, our proposed ALNet also achieved satisfying results. The proposed method showed a remarkable improvement in the work done previously.