A hierarchical coarse-to-fine perception for small-target categorization of butterflies under complex backgrounds

Article type: Research Article

Authors: Lin, Zhongqi^{a; b} | Jia, Jingdun^b | Gao, Wanlin^{a; b; *} | Huang, Feng^{b; c; *}

Affiliations: [a] College of Information and Electrical Engineering, China Agricultural University, Beijing, China | [b] Key Laboratory of Agricultural Informatization Standardization, Ministry of Agriculture and Rural Affairs, Beijing, China | [c] College of Science, China Agricultural University, Beijing, China

Correspondence: [*] Corresponding author. Wanlin Gao, E-mail: gaowlin@cau.edu.cn and Feng Huang, E-mail: huangfeng@cau.edu.cn.

Abstract: Small-target categorization of butterflies suffers from large-scale search space of candidate target locations, subtler discriminations, camouflaged appearances, and complex backgrounds. Precise localization and domain-specific discrimination extraction are crucial for this issue. In this work, a novel hierarchical coarse-to-fine convolutional neural network (C-t-FCNN) was proposed. It consists of CoarseNet and FineNet, which incorporate object-level and part-level representations into framework. Specifically, the coarse-grained features containing the orientation description are generated by CoarseNet, while the fine-grained discriminations with semantic distinctiveness are captured by FineNet. Next, the correspondences are established to mark the target regions, background regions, and mismatched regions depending on the quantification of scale-invariant feature transform (SIFT) descriptors. Then, the features are subsampled via spatial pyramid pooling (SPP) for size uniformity and integration. Finally, the irrelevant background and mismatched regions are eliminated by the support vector machine (SVM) with a radial basis function (RBF) kernel, leaving only the target-specific patches for finer-scale extraction. Hence the numeracy can be economized from identifying irrelevant areas and can be rescheduled in feature extraction and final decision, which can suppress time complexity simultaneously. A total of 119,016 augmented butterfly images spanning 47 categories are utilized for model training, while 13,734 images are evaluated for effectiveness verification. The C-t-FCNN delivers impressive performance, i.e., it achieves a validation accuracy of 92.08% and a testing accuracy of 91.6%, which outperforms state-of-the-arts.

Keywords: Convolutional neural network, coarse-to-fine perception, deep learning, small target categorization, butterfly images

DOI: 10.3233/JIFS-190747

Journal: Journal of Intelligent & Fuzzy Systems, vol. 38, no. 3, pp. 3463-3487, 2020