目標導向區塊強化機制之少樣本影像分類;TAP-Net: Target-Aware Patch Enhancement Mechanism for Few-Shot Image Classification

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/98351

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98351

Title:	目標導向區塊強化機制之少樣本影像分類;TAP-Net: Target-Aware Patch Enhancement Mechanism for Few-Shot Image Classification
Authors:	陳詩云;Chen, Shih-Yun
Contributors:	資訊工程學系
Keywords:	少樣本影像分類;YOLO;特徵增強;Few-Shot Image Classification;YOLO;Feature Enhancement
Date:	2025-07-28
Issue Date:	2025-10-17 12:40:15 (UTC+8)
Publisher:	國立中央大學
Abstract:	在基於局部特徵的少樣本影像分類任務中，模型常因區域類別相似度高而導致分類錯誤。因此，我們提出一種整合目標導向區域強化機制的少樣本分類架構 TAP-Net。該方法結合 YOLO 偵測模型，定位圖片中的目標區域，並設計兩項模組：目標導向區塊重加權（Target-Guided Patch Reweighting）機制與目標導向特徵增強機制（Target-Guided Feature Enhancement），分別強化模型對重要區域的權重與特徵表示。此外，本研究亦提出一種目標導向的資料增強策略，透過模糊非目標區域以減少背景資訊，進一步提升模型泛化能力。在 miniImageNet 資料集上的實驗結果顯示，我們的方法在 5-way 1-shot 與 5-way 5-shot 任務下的分類準確率分別為 71.31% 與 84.99%，皆優於原始 FewTURE 架構，證明本研究所提出方法能有效提升模型的泛化能力，為少樣本分類任務提供了一種新的解決方法。 ;In patch-based few-shot image classification, models often suffer from misclassification due to high inter-class similarity within local regions. To address this challenge, we propose TAP-Net, a novel few-shot classification framework that incorporates a target-guided region enhancement mechanism. Our method leverages the YOLO detection model to localize target regions within images and introduces two key modules: Target-Guided Patch Reweighting, which emphasizes the importance of semantically relevant patches, and Target-Guided Feature Enhancement, which strengthens the representation of crucial areas by enriching their contextual features. Additionally, we propose a target-guided data augmentation strategy that reduces background interference by applying Gaussian blur to non-target regions, thereby improving the model’s generalization capability. Experimental results on the miniImageNet dataset demonstrate that TAP-Net achieves classification accuracies of 71.31% for the 5-way 1-shot task and 84.99% for the 5-way 5-shot task—both surpassing the performance of the original FewTURE framework. These findings validate the effectiveness of our proposed approach in enhancing feature discriminability and model generalization, offering a new direction for few-shot classification research.
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	3	View/Open

社群 sharing

Loading...