視障人士自主生活與戶外行動的權利，受到感官能力的限制，對視障者而言是一個難以達成的任務。即便配合視覺輔具與無障礙措施，獲取行動的助益，但招牌的獨特性圖騰設計，傳統機器視覺較不易辨別。為此本研究針對街景商店影像，提出一個創新的招牌偵測與檢索系統，我們以神經網路進行偵測，找出街景上多個招牌物件位置，藉由招牌影像建立一個以圖像資料為索引的結構，使用影像均值雜湊、感知雜湊及差異雜湊算法，可以處理商業獨特性圖騰的識別及無數種商店招牌的種類，並結合VP決策樹迅速搜索的優勢，以檢索方式尋找最相似的特徵節點。街景招牌資料集使用自行蒐整建立的影像，系統的招牌定位模組召回率達為84%，而檢索模組Rank1及Rank5都能成功檢索命中，最後使用偵測與檢索整體實驗平均精確度達86%。本系統開發提供給視障者的視覺輔助，回饋當前店家的類型資訊，使視障朋友也能與常人一樣感知與行動決策。;The rights of the visually impaired to live independently and to move outdoors are limited by their perception abilities, which is a difficult task for the visually impaired. Even with visual aids and barrier-free measures to help them living better, the unique totem design of the signboard is not easy to distinguish by traditional machine vision. Therefore, in this article we propose an innovative signboard detection and retrieval system for street view store images. We use neural network to detect multiple signboard object positions on the street view, and build a structure indexed by image data from the signboard images. Image mean hashing, perceptual hashing and difference hashing algorithms can handle the recognition of business unique totems and countless types of store signs, and combine the advantages of rapid search of the VP decision tree method to find the most similar feature matrix by retrieval. The Street View signboard dataset uses images created by self-searching. The system’s signboard positioning module recall rate reached 84%, and the retrieval modules Rank1 and Rank5 can successfully retrieve hits. Finally, the average accuracy of the overall experiment of detection and retrieval is up to 86%. This system is developed to provide visual aids for the visually impaired, and feedback the current store type information, so that the visually impaired friends can perceive and make decisions like normal people.
