圖像資料結構化與分類的探討

Abstract

資料以各種形態存在於我們生活中，寫過的每一篇文章，甚至拍過的每一張照片，透過適當的數位化皆可由量化分析挖掘出其中的重要訊息。過去資料分析大多侷限在數字格式，隨著電腦相關技術的發展，資訊解讀擴展至文字、圖像、音樂等各種類型的資料，我們的生活因為資訊傳遞快速、即時判讀而更加便利，影像辨識、自動駕駛等應用就是眾所周知的應用。資料格式多元、傳遞交換便捷，都是大數據時代的特點，使得資訊安全及品質愈形重要，如何解讀龐雜的大數據，更是政府及個人必備的關鍵能力。不具固定格式資料稱為非結構資料，而解讀這類型資料的首要挑戰為數位化格式，但轉檔方式與研究目標、資料屬性關係密切，很難訂出一個絕對標準。\n以圖像辨識為例，資料應轉換成三原色（紅綠藍，RGB：Red、Green、Blue）或是圖像形狀及大小，至今仍無定論；即便是以顏色紀錄，是否也需考量色彩飽和度、亮度等資訊？有鑑於圖像資料尚無統一的格式化，本文以視覺感受的方式定義變數，比較冷暖色、RGB、灰階、邊緣檢測、分割圖像等方法，協助分類不同風格的圖像。由於圖像辨識結果多半與其屬性有關（Data Dependent），本文分析三種類型的圖像資料：臺灣報紙頭版、美國Vogue雜誌封面、十九世紀油畫（現實派、印象派），其內容包含文字、照片（及圖片）、繪畫，再結合統計分析、機器學習模型，藉由電腦模擬與交叉驗證，探討不同的圖像格式方法、模型及其參數的分類準確性。研究結果顯示，分類準確性隨著解析度上升而提高，100100即有不錯的效果；而圖像格式化中以分割圖像法最佳，用於不同圖像資料及分析模型的分類準確性較高。

Data exists in various forms, including articles we write and photos we take. We can dig out important information through appropriate digitization and quantitative analysis. Data analysis was usually limited to data with digital format and now has expanded to all kinds of data, such as text, images, and music. Rapid data transmission and real-time analysis brings convenience to our lives, and image recognition and autonomous driving are two famous applications. The ability of analyzing messy and complicated data is a key ability in Big Data era. However, more than 90% of data are unstructured data and it needs to convert them into digital format before plugging into data analysis. But the conversion method is closely related to the research objectives and data attributes. Taking image recognition as an example, there is no consensus if the image data should be converted into three primary colors (red, green and blue, RGB) or their shape and size should also be considered.\nThis study aims to explore different structurization methods for image data and evaluate which method, after plugging into classification models, has the highest accuracy in classifying images . We consider three types of image data: Taiwanese newspapers, Vogue magazine, and nineteenth century oil paintings, since the classification results are often data-dependent. We will apply statistical and machine learning models to explore classification accuracy of different image format methods. The analysis results show that the classification accuracy increases when the resolution becomes higher, and 100100 resolution can provide sufficiently satisfactory results. We found that the splitting method has highest accuracy in image classification, for three types of image data and different classification models.