Learning Deep Bilinear Transformation for Fine-grained Image Representation

By Heliang Zheng, Jianlong Fu, Zhengjun Zha, Jiebo Luo. 2019

178 citations recorded by the source corpus.

Bilinear feature transformation has shown the state-of-the-art performance in learning fine-grained image representations. However, the computational cost to learn pairwise interactions between deep feature channels is prohibitively expensive, which restricts this powerful transformation to be used in deep neural networks. In this paper, we propose a deep bilinear transformation (DBT) block, which can be deeply stacked in convolutional neural networks to learn fine-grained image representations. The DBT block can uniformly divide input channels into several semantic groups. As bilinear transformation can be represented by calculating pairwise interactions within each group, the computational cost can be heavily relieved. The output of each block is further obtained by aggregating intra-group bilinear features, with residuals from the entire input features. We found that the proposed network achieves new state-of-the-art in several fine-grained image recognition benchmarks, including CUB-Bird, Stanford-Car, and FGVC-Aircraft.

Subfields: Artificial Intelligence, Biomedical Engineering, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Biomedical Engineering

In Computer Vision and Pattern Recognition

Citations in the corpus

Self-Sustaining Representation Expansion for Non-Exemplar Class-Incremental Learning (2022, neutral): “To retain the representation [42] of the old class and guarantee the unbiased training of the new class, we propose a dynamic structure reorganization strategy.”
Fine-Grained Image Analysis With Deep Learning: A Survey (2021, methodology): “[105] applied bilinear pooling to feature channel groups where the bilinear transformation is represented by calculating pairwise interactions within each group.”
Semantics-Guided Neural Networks for Efficient Skeleton-Based Human Action Recognition (2019, background): “, machine translation [45] and image recognition [62].”
Document-Level Relation Extraction with Adaptive Thresholding and Localized Context Pooling (2020, methodology): “To reduce the number of parameters in the bilinear classiﬁer, we use the group bilinear (Zheng et al., 2019), which splits the embedding dimensions into k equal-sized groups (Tang et al., 2020b) and applies bilinear within the groups: where W ir ∈ R d/k × d/k for i = 1 ...k are model parameters, P…”
TransFG: A Transformer Architecture for Fine-grained Recognition (2021, methodology): “[53] adopted the idea of group convolution to first split channels into different groups by their semantic meanings and then do the bilinear pooling within each group without changing the dimension thus it can be integrated into any existed backbones directly.”