Multi-Proxy Loss:基於度量學習提出之損失函數用於細粒度圖像檢索;Multi-Proxy Loss: For Deep Metric Learning on Fine-grained Image Retrieval

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/89784

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/89784

Title:	Multi-Proxy Loss:基於度量學習提出之損失函數用於細粒度圖像檢索;Multi-Proxy Loss: For Deep Metric Learning on Fine-grained Image Retrieval
Authors:	林宛儀;Lin, Wan Yi
Contributors:	資訊工程學系
Keywords:	度量學習;距離學習;圖像檢索;細粒度圖像;卷積神經網路;Deep Metric Learning;Distance metric learning;Image Retrieval;Fine-grained;Convention Network
Date:	2022-07-21
Issue Date:	2022-10-04 11:59:38 (UTC+8)
Publisher:	國立中央大學
Abstract:	本篇論文針對圖像檢索(Image retrieval )的任務上提出了一個新的損失函數。此方法基於Proxy_NCA以及Proxy_Anchor的方法上加上了多個代表點的方法，來提升樣本的豐富性。使得Batch size減少的情況下也能達到跟原來較大的batch size一樣的效果。並且使用SoftMax函數對類內代表點做加權。使得重要的代表點能得到更多的學習資源。除此損失函數地改良之外，也對現有的ResNet50進行了修改，只使用RestNet50的前三層做為特徵擷取，取消了ResNet50第三層的下採樣。並且加入了Attention機制取代原本ResNet50的第四層。Attention使用了SoftPlus函數對特徵圖的特徵做加權。使得重要的特徵能更明顯，不重要的特徵減少關注度。相較於傳統Attention使用SoftMax函數能得到更好的效果。不管是新提出的損失函數，或是改良過後的ResNet50都相較於原始方法Recall@1都有很大的提升。;In this paper, we propose a new loss function for Image Retrieval task. The new loss function makes an improvement based on Proxy-NCA and Proxy-Anchor Loss by adopting multiple proxies, to promote positive sample variety. Its shows better performance than Proxy-Anchor Loss even in the small batch size. Besides, we weighted intra-class proxy by SoftMax function to make important samples receive a higher gradient while training. In addition, we make some changes on ResNet50 by only using the first three-layer and adding a new attention module by using SoftPlus function to replace SoftMax. Finally, we obtain well results on recall@1 via our new method.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	46	View/Open

社群 sharing

Loading...