Deep Large-Margin Rank Loss for Multi-Label Image Classification

The large-margin technique has served as the foundation of several successful theoretical and empirical results in multi-label image classification. However, most techniques are only suitable to shallow models with preset feature representations a few neural networks enforce margins at output layer, which not well for deep networks. Based on technique, rank loss function any network structure is proposed, able impose margin chosen set layers network, allows choosing ℓp norm (p≥1) metric measuring between labels applicable architecture. Although complete computation O(C2) time complexity, where C denotes size label set, would cause scalability issues when large, negative sampling was proposed make scale linearly C. Experimental two large-scale datasets, VOC2007 MS-COCO, show that ranking improves robustness model classification tasks while enhancing model’s anti-noise performance.