Proceedings of the
The Nineteenth International Conference on Computational Intelligence and Security (CIS 2023)
December 1 – 4, 2023, Haikou, China
Bilateral Branch Network Framework based on Multi-View Feature Aggregation
Computer Science and Technology, Xidian University, China.
ABSTRACT
The problem of natural visual recognition with long-tailed data distribution presents significant challenges, as traditional classifiers tend to be biased towards the majority or `head' classes, often leading to inaccurate predictions. To address this issue, this paper proposes a Bilateral Branch Network Framework based on Multi-View Feature Aggregation (BMVFA) specifically designed for the classification of imbalanced image data. The proposed framework enhances the discriminative capacity of features by aggregating multi-view representations of each sample, thereby amplifying the feature differences between disparate classes. Additionally, the model incorporates a residual accumulation mechanism at its terminal layer to mitigate classifier bias introduced by data imbalance. A comprehensive set of experiments on four long-tailed visual recognition datasets is presented, alongside comparisons with state-of-the-art algorithms. The empirical results substantiate the superior performance of our proposed network in handling long-tailed data recognition tasks.
Keywords: Long-tailed data, Image classification, Feature enhancement.
Download PDF
Computer Science and Technology, Xidian University, China.
ABSTRACT
The problem of natural visual recognition with long-tailed data distribution presents significant challenges, as traditional classifiers tend to be biased towards the majority or `head' classes, often leading to inaccurate predictions. To address this issue, this paper proposes a Bilateral Branch Network Framework based on Multi-View Feature Aggregation (BMVFA) specifically designed for the classification of imbalanced image data. The proposed framework enhances the discriminative capacity of features by aggregating multi-view representations of each sample, thereby amplifying the feature differences between disparate classes. Additionally, the model incorporates a residual accumulation mechanism at its terminal layer to mitigate classifier bias introduced by data imbalance. A comprehensive set of experiments on four long-tailed visual recognition datasets is presented, alongside comparisons with state-of-the-art algorithms. The empirical results substantiate the superior performance of our proposed network in handling long-tailed data recognition tasks.
Keywords: Long-tailed data, Image classification, Feature enhancement.
Download PDF