Open AccessAttention-Based Dual-Branch Cascade Network for Multi-Label Image Recognition
Author(s) -
Xingyu Li
Publication year - 2022
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/2224/1/012018
Subject(s) - pascal (unit) , computer science , artificial intelligence , pattern recognition (psychology) , cascade , dual (grammatical number) , benchmark (surveying) , representation (politics) , visualization , image (mathematics) , feature (linguistics) , key (lock) , object (grammar) , task (project management) , machine learning , management , art , linguistics , chemistry , philosophy , literature , computer security , geodesy , chromatography , politics , political science , law , economics , programming language , geography
Multi-label image recognition is a practical task which aims to predict all concerned objects in an image, and the correlation between labels is considered as the key to solve multilabel problem. Previous approaches focus on exploring and exploiting the underlying relations between labels, by which the performance of most labels is improved significantly. But small object and label tail problems are ignored, and negatively influences the overall results. For ameliorating above problems, we propose a unified deep neural network named Attention-Based Dual-Branch Cascade Network (ADC Net), which contains Main Branch and Auxiliary Branch. ADC Net cascades to predict labels, and obtains the final result by element-wise adding. Attention modules in each branch contribute to recognizing small objects. Top-Down-Attention Module (TDAM) utilizes the preceding prediction map to guide the Auxiliary Branch. Besides, expanding training dataset is applied for learning the feature representation of tail labels. Our proposed methods are evaluated on two benchmark datasets: MS-COCO and VOC PASCAL 2007 datasets, and achieves state-of-the-art performance. Results of small objects, tail labels and visualization also prove the effectiveness of our method.