Abstract | ||
---|---|---|
•We realize distorted target recognition by presenting a CapsNetSIFT architecture (cf. Fig. 1). It can simultaneously enable accurate positioning of interest regions and comprehensive learning of discriminative features whilst boasting invariance to visual distortions.•We propose a parallel MD-CapsNet in CapsNetSIFT for boosting representability. The customized structure, hyperparameter setting, and dynamic routing agreement have been demonstrated.•We propose a VM-SIFT in CapsNetSIFT to establish correspondence among capsule encoding vectors of standard images and distorted ones. This may represent the first attempt to realize feature matching of capsule vectors.•We conduct quantitative experiments of distorted target categorization on four benchmarks (CUB-200-2011, Stanford Dogs, Stanford Cars, and our hand-crafted rice growth dataset). Evaluation results reveal our advantage over state-of-the-arts. |
Year | DOI | Venue |
---|---|---|
2021 | 10.1016/j.neucom.2021.08.087 | Neurocomputing |
Keywords | DocType | Volume |
Capsule network (CapsNet),Scale-invariant feature transform (SIFT),Distorted target categorization,Feature matching | Journal | 464 |
ISSN | Citations | PageRank |
0925-2312 | 0 | 0.34 |
References | Authors | |
0 | 4 |
Name | Order | Citations | PageRank |
---|---|---|---|
Zhongqi Lin | 1 | 3 | 3.12 |
Wanlin Gao | 2 | 6 | 7.58 |
Jingdun Jia | 3 | 3 | 2.44 |
Feng Huang | 4 | 1 | 1.38 |