基於卷積神經網路之文字來源印表機鑑識

Abstract

近年來隨著資訊科技的蓬勃發展與多媒體設備技術的提昇，不論是工作、學習或日常生活中的各種活動，都和數位內容息息相關。而數位影像處理的技術，也被應用在許多領域，範圍包括醫學、數學、生物學、氣象學、工程科學等領域。其中，最普遍的數位影像的輸出裝置不外乎是印表機。然而，利用印表機從事偽造變造貨幣、偽造文件、侵害著作權…等犯罪不斷地出現在新聞媒體中。在現行的環境下，一般對於數位影像輸出的文件皆是由專業人員進行數位影像來源的辨識。儘管如此，鑑於辨識技術的迅速發展，對於數位內容來源識別，期望提供更有效率的技術來節省時間與人力成本。 本論文針對從印表機列印出的文件，擷取常用的文字與圖片，進行前置影像處理並擷取特徵，再進一步利用卷積神經網路（Convolutional Neural Network，CNN），進行印表機設備來源之識別。在本篇論文的實驗中，透過利用卷積神經網路（Convolutional Neural Network，CNN）方法，取代在特徵擷取、選擇等資料預處理的過程常是需要有人工參與的作業。並使用訓練數據進行測試，最佳實驗測試結果準確度可達99%。 本研究的實驗結果顯示，採用所提出之掃描圖像的分析技術，其辨識準確度及執行時間顯著的縮短皆優於既存方法，因此可以證明本研究中所提出的方法可以實際應用於列印文件來源之印表機鑑識。 關鍵字：影像檢索、深度學習、卷積神經網路、裝置辨識、特徵選擇

Along with the speedy development of information technology and the improvement of multimedia equipment in recent years, digital content is closely related to works, studies, and various activities in daily life. Digital image processing technology has also been applied to various fields like Medical Science, Mathematics, Biology, Meteorology, Engineer Science, and so on. Among those applications, the printer is the most commonly used device for digital content output. However, the crimes engaged in forged currency, altered documents, and encroachment of copyright by utilizing printers emerges illegally. Generally, the source of digital image output file is identified by the professional experts currently. Despite the mentioned fact, given the rapid development of identification technology, it is expected to provide more efficient methods of identifying the printed source to save time and labor cost.

In this study, the commonly used texts and pictures are extracted from the printed documents. Extracted text and picture will go through the pre-image process to retrieve its features. The identification experiments of printer source will be executed by utilizing Convolutional Neural Network (CNN). In this paper, the accuracy ratios of the best experimental results can reach up to 99% by utilizing CNN approach with least manual participation which usually involves feature extraction, selection, and other data preprocessing procedures.

The experimental results of this study indicate that the proposed image analysis technology is superior to the existing approaches in terms of the accuracy of identification with short execution time. Based on the previous mentioned facts, our approach has proved that it can be applied to the printer source identification with low complexity.

Keywords：Image Retrieval, Deep Learning, Convolutional Neural Network, Device Identification, Feature Selection