請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92870
標題: | 邱商標:中文商標偵測與辨識 ChiuLogo: Chinese Logo Detection and Recognition |
作者: | 邱議禾 I-Ho Chiu |
指導教授: | 傅楸善 Chiou-Shann Fuh |
關鍵字: | 開放集商標識別,商標檢測,度量學習,場景文字識別, Open-Set Logo Recognition,Logo Detection,Metric Learning,CLIP,YOLO,Scene Text Recognition, |
出版年 : | 2024 |
學位: | 碩士 |
摘要: | 本論文提出了一種名為ChiuLogo邱商標的方法,解決了計算機視覺中標誌檢測和識別的複雜挑戰,特別是在多樣化標誌的開放集標誌識別領域。邱商標利用先進的深度學習技術,整合了CN-CLIP (Chinese-Contrastive Language-Image Pretraining) 模型進行識別和YOLOv8 (You Only Look Once) 框架進行檢測,創建了一個強大的解決方案。邱商標的一個獨特特點是融合了來自CN-CLIP和場景文字識別模型PARSeq (Permuted Auto-Regressive Sequence modeling) 的雙重編碼器,這些編碼器通過具有創新性的Proxy Anchor Loss的度量學習進一步增強。這種整合有效地處理了標誌的圖形和文字組件。此外,邱商標利用Tip-Adapter方法優化識別性能,導致準確性和適應性顯著提高。在LogoDet-3K和QMUL-OpenLogo等基準數據集上的廣泛測試展示了邱商標實現了state-of-the-art的結果,將標誌檢測和識別系統的能力推向了新的高度。 This thesis presents ChiuLogo, a method that addresses the intricate challenges of logo detection and recognition within computer vision, specifically focusing on open-set logo recognition among a varied array of logos. ChiuLogo harnesses advanced deep learning techniques, integrating the CN-CLIP model for recognition and the YOLOv8 framework for detection to create a robust solution. A distinctive feature of ChiuLogo is the fusion of dual embeddings from CN-CLIP and the Scene Text Recognition model PARSeq, which are further enhanced through metric learning with the innovative Proxy Anchor Loss. This integration effectively processes both graphical and textual components of logos. Additionally, ChiuLogo utilizes the Tip-Adapter method to refine recognition performance, leading to significant improvements in accuracy and adaptability. Extensive testing on benchmark datasets, notably LogoDet-3K and QMUL-OpenLogo, showcases ChiuLogo''s ability to achieve state-of-the-art results, propelling the capabilities of logo detection and recognition systems to new heights. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92870 |
DOI: | 10.6342/NTU202401255 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-112-2.pdf | 16.24 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。