Long-tailed image recognition through balancing discriminant quality

Yan-Xue Wu,Fan Min,Ben-Wen Zhang,Xian-Jie Wang

DOI: https://doi.org/10.1007/s10462-023-10544-x

IF: 9.588

2023-07-07

Artificial Intelligence Review

Abstract:Long-tailed image recognition is a challenging task in real scenes with large-scale data. Popular strategies, such as loss reweighting and data resampling, aim to reduce the model bias toward head classes. Specifically, different loss reweighting approaches explore various endogenous or exogenous measures. In this paper, we study a new endogenous measure called discriminant quality (DQ) by considering validation accuracy and discriminant uncertainty. DQ takes advantage of continuous information over a period of time. It is more robust than instantaneous information because of the mitigation of measuring instability caused by random perturbations during training. Additionally, the weight of each class is automatically rebalanced based on DQ. Consequently, the class weight supports the design of a dynamic updating strategy for the significance of the DQ difference. Experiments on MNIST-LT, CIFAR-100-LT, ImageNet-LT, and Places-LT demonstrated the superiority of DQ over state-of-the-art ones in terms of prediction accuracy.

computer science, artificial intelligence

What problem does this paper attempt to address?

### What problem does this paper attempt to solve?

This paper primarily investigates the challenging problem of long-tailed distribution image recognition. Specifically, it explores a novel endogenous metric—Discriminant Quality (DQ), aiming to reduce the model's bias towards head categories by balancing the discriminant quality of different categories.

#### Main Contributions:
1. **Proposing the DQ Metric**: Defines DQ by combining validation accuracy and discriminant uncertainty to evaluate the training quality of each category.
2. **Utilization of Temporal Continuous Information**: Uses information over continuous time periods rather than instantaneous information to measure DQ, enhancing robustness.
3. **Dynamic Update Strategy**: Designs a dynamic update strategy to adjust the importance of DQ differences, thereby achieving automatic rebalancing of category weights.
4. **Loss Function Design**: Proposes three different loss functions based on accuracy, uncertainty, and DQ to rebalance category weights.

#### Experimental Results:
Experiments on multiple datasets (such as MNIST-LT, CIFAR-100-LT, ImageNet-LT, and Places-LT) show that this method outperforms existing advanced methods in terms of prediction accuracy. Specifically, compared to the standard cross-entropy loss, this method significantly improves average performance under different imbalance factors.