延展式區域三元化圖形特徵描述子之比例式降維法

Abstract

區域二/三元化樣式與其各種變型被廣泛應用於物件辨識中的特徵描述，然而現有的區域特徵描述方式，普遍存在適用時機的問題，也就是針對不同類型的圖像資料庫，必須選用符合該圖片性質的描述法，方能達到較佳的辨識效果，舉例而言，處理材質影像時多使用uniform pattern，而進行人臉偵測或表情辨識時則多採用一般型的區域二/三元化樣式。\n\n本論文的目標是建構一個通用型的區域三元化樣式，使其一體適用於各類圖型辨識的任務，我們以延展式區域三元化樣式(Extended local ternary patterns, ELTP)為基礎，探討各種降維演算法的結合機制，並提出可行的樣式定義方法，我們針對ETLP 中的uniform pattern定義重新思考，藉由大規模實驗與統計，探討各類uniform pattern 的從屬關係與出現比例，並依據比例原則，在降維階段分配適當之維度，稱之為比例式降維法。\n\n本論文針對比例式降維後的ELTP之抗噪性、描述力與通用性進行深度的分析與廣泛的實驗，以驗證此類圖像描述方法之效能。此外，由於特徵樣式的編碼在某些情況下，容易受到雜訊影響而產生變化，因此我們亦提出了結合模糊理論的方式加以改善。以上改良方式之效果，都在各式實驗中獲得驗證。

Local binary/ternary pattern and its derivatives have been widely employed to represent low-level features in many pattern recognition tasks. However, existing local descriptors fail to achieve universal applicability in the sense that specific types of local binary patterns are better suited for certain collections of images. For example, uniform local binary patterns are preferred when dealing with textures, while regular local binary/ternary patterns are adopted for face detection and facial expression recognition.\n\nThis thesis proposes a universally applicable local descriptor based on the extended local ternary pattern (ELTP) to address the above concern. We exploit the feasibility of combining dimensionality reduction techniques to derive a novel local descriptor that is suitable for all kinds of object recognition applications. Specifically, we investigate all possible definitions of uniform patterns under ternary encoding scheme and study their properties. This enables us to devise a dimensionality assignment algorithm in which the allocated dimension is proportional to the appearance rate of the corresponding pattern group.\n\nThe newly defined extend local ternary pattern using commensurate dimensionality reduction (ELTP-CDR) technique has been extensively tested to analyze its universality, discriminability, and noise sensitivity. To further enhance the robustness of this class of local descriptor, we also incorporate fuzzy logic to derive fuzzy representations of the original ELTP. The efficacy of the proposed methods has been validated through several experiments targeted at texture recognition tasks.