Abstract:Coarse-grained classification and classification at different hierarchy levels are important in many engineering applications. Conventional approaches include post-inference aggregation of fine-grained predictions or multi-task learning with its disadvantages: necessity to incorporate post-prediction mechanisms, possible mistakes between unrelated categories when generic fine-grained models are used, and the need of training multi-task models with joint procedures. To overcome them, this paper proposes a novel method for extracting coarse-grained classifiers from large Convolutional Neural Networks (CNNs) trained on fine-grained tasks with no additional training. Leveraging the knowledge transfer capabilities of CNNs and acknowledging that a model proficient in a challenging task should inherently address easier, hierarchically related problems, the approach finds and aggregates available representatives of coarse-grained classes within existing fine-grained models. Unlike conventional approaches, the proposed technique aggregates the fine-grained model's weights and not its predictions. As a result, a new coarse-grained model is obtained with the same feature extractor as the original model. The original and new models can be used to create a multi-output model without the multi-task learning. The method eliminates additional post-prediction mechanisms and limits the mistakes between unrelated categories. It can be used with three strategies to find the sets of fine-grained classes for weight aggregation. They are based on external WordNet knowledge or internal network knowledge. Two of the methods leverage the concepts of semantic and visual similarity. Experimental results with MobileNetV2, InceptionV3, ResNet50V2, EfficientNetV2B0 and ConvNeXt-Small trained on ImageNet demonstrate up to 98% accuracy in the coarse-grained classification, which is superior to the post-prediction aggregation.

Multi-task Convolutional Neural Network with Coarse-to-fine Knowledge Transfer for Long-Tailed Classification

Hierarchical Convolutional Neural Network with Knowledge Complementation for Long-Tailed Classification

Hierarchical long-tailed classification based on multi-granularity knowledge transfer driven by multi-scale feature fusion

Multiscale 3-D-2-D Mixed CNN and Lightweight Attention-Free Transformer for Hyperspectral and LiDAR Classification

Coarse-to-fine Trained Multi-Scale Convolutional Neural Networks for Image Classification.

Cumulative Dual-Branch Network Framework for Long-Tailed Multi-Class Classification

Balanced Classification: A Unified Framework for Long-Tailed Object Detection

Extracting coarse-grained classifiers from large Convolutional Neural Networks

Long-Tailed Recognition by Hierarchical Rebalancing Dual-Classifier

Mixed Mutual Transfer for Long-Tailed Image Classification

Dynamic collaborative learning with heterogeneous knowledge transfer for long-tailed visual recognition

Relieving the Incompatibility of Network Representation and Classification for Long-Tailed Data Distribution

Pay Attention to Convolution Filters: Towards Fast and Accurate Fine-Grained Transfer Learning

Long-Tailed Visual Recognition via Self-Heterogeneous Integration with Knowledge Excavation

One‐stage self‐distillation guided knowledge transfer for long‐tailed visual recognition

On Size-Oriented Long-Tailed Graph Classification of Graph Neural Networks

Calibrating Class Activation Maps for Long-Tailed Visual Recognition

Two-Stage Fine-Tuning: A Novel Strategy for Learning Class-Imbalanced Data

Feature Fusion Network for Long-Tailed Visual Recognition

HFT-ONLSTM: Hierarchical and Fine-Tuning Multi-label Text Classification

Improving Long-Tailed Classification from Instance Level