Naming conventions-based multi-label and multi-task learning for fine-grained classification

Abstract

This paper proposes a fine-grained image classification architecture using multi-task learning. The structure of the fine-grained classification network uses ResNest as the feature extraction layer of the multi-task hard parameter sharing mode with the fine-grained category label regression branch based on multi-hot naming conventions and classification branch based on cross-entropy loss with one-hot encoding. The coupling between the two branches enables multi-task classification through hyperparameter weighting. Subsequently, comparison and ablation experiments were performed on the public datasets of Stanford Cars, CUB-200-2011 and FGVC-Aircraft. The experimental result shows multi-label regression, multi-task learning and label smoothing can effectively improve the generalization of the model and increase the inter-class distance of the previous layer at the network output terminal, and reduces the intra-class distance.