A Deep Model for Partial Multi-label Image Classification with Curriculum-based Disambiguation

作     者：Feng Sun Ming-Kun Xie Sheng-Jun Huang Feng Sun;Ming-Kun Xie;Sheng-Jun Huang

作者机构：MIIT Key Laboratory of Pattern Analysis and Machine IntelligenceCollege of Computer Science and TechnologyNanjing University of AeronauticsandAstronauticsNanjing211106China 

出 版 物：《Machine Intelligence Research》 (机器智能研究（英文版）)

年 卷 期：2024年第21卷第4期

页      面：801-814页

核心收录：

学科分类：08[工学] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Key Research and Development Program of China, NKRDPC, (2020AAA0107000) National Key Research and Development Program of China, NKRDPC National Natural Science Foundation of China, NSFC, (62222605) National Natural Science Foundation of China, NSFC Natural Science Foundation of Jiangsu Province, (BK20222012, BK20211517) Natural Science Foundation of Jiangsu Province 

主　　题：Partial multi-label image classification curriculum-based disambiguation consistency regularization label difficulty candidatelabel set. 

摘      要：In this paper,we study the partial multi-label(PML)image classification problem,where each image is annotated with a candidate label set consisting of multiple relevant labels and other noisy *** PML methods typically design a disambiguation strategy to filter out noisy labels by utilizing prior knowledge with extra assumptions,which unfortunately is unavailable in many real ***,because the objective function for disambiguation is usually elaborately designed on the whole training set,it can hardly be optimized in a deep model with stochastic gradient descent(SGD)on *** this paper,for the first time,we propose a deep model for PML to enhance the representation and discrimination *** the one hand,we propose a novel curriculum-based disambiguation strategy to progressively identify ground-truth labels by incorporating the varied difficulties of different *** the other hand,consistency regularization is introduced for model training to balance fitting identified easy labels and exploiting potential relevant *** experimental results on the commonly used benchmark datasets show that the proposed method significantlyoutperforms the SOTA methods.