基於圖像內容特徵與決策樹之麵包辨識系統

Abstract

現今市場上大多數商品結帳系統都採用掃描條碼(barcode)的方式來進行，然而針對生鮮食品諸如肉品、海鮮及蔬菜等，無法輕易貼上條碼之產品，往往需要增添額外人工流程來協助結帳。然而麵包因新鮮出爐具有高溫，包裝易有水氣，為求提升賣相之考量，通常不會額外包裝，也就無法輕易貼上條碼。因此透過影像處理是一個很好的解決方案，利用相機直接擷取麵包影像，經由辨識系統來識別產品，不僅能保存麵包原始樣貌，更可以提高結帳流程效率。 因此本論文提出一個基於圖像內容特徵與決策樹分類的麵包辨識系統，透過數位相機或視訊攝影機鏡頭拍攝麵包，以影像處理方式設計系統，希望藉由自動化流程，來提升目前人工結帳效率。本研究主要由三大流程所構成：麵包影像前處理、圖像內容特徵擷取和決策樹分類演算法。首先透過影像前處理來進行前景(麵包)與背景分離，處理成二值化影像，並擷取影像上的圖像內容特徵，包含基本幾何色彩特徵與轉折點偵測等形狀特徵，後續統計特徵極值範圍，進行範圍比對以切割分區，並搭配信息熵理論，計算熵值(Entropy)挑選最佳屬性，以產生決策樹分類樣式。經由實驗顯示本研究針對48種麵包進行分類測試(每種以30個樣本進行訓練，以10個樣本進行測試)，結果具備有93.75%的辨識率，可提升結帳效率並降低人力資源成本。未來此架構流程可應用於物件辨識(Object Recognition)領域。

Nowadays most merchandise checkout systems are used by way of scanning each items barcode. However, fresh food such as meat, seafood, bread, and vegetables cannot be scanned as easily since they usually do not come with a barcode. The handling of these products often require a clerk’s assistance. For example, fresh bread that is being cooled cannot be stored in the plastic bags right away, therefore making it difficult to attach a barcode label to the item. As a result, an image processing aided system would be a suitable solution to assist with issues related to barcode labeling, while also helping to improve overall checkout efficiency. In this thesis, we propose an intelligent bread recognition system (IBRS) by using photographic equipment to capture bread images. The system includes three main modules: original image preprocessing, content-based feature extraction and decision tree classification. The input images will be first segmented and processed into binary images. Then, the geometry, color, and shape feature will be extracted from the preprocessed images. After that, the extreme range of features will be analyzed, and the best attribute will be chosen through calculating the entropy. Finally, the system will generate a decision tree classification model, which will then be used to identify the input bread images. The experiment indicates that the classification for 48 kinds of bread samples resulted in 93.75% accuracy. Each test used 30 samples for training and 10 samples for testing. As a result, our system can enhance checkout efficiency and reduce labor costs. In the future, the process architecture can be applied to Objection Recognition field.