生物啟發之色彩與質感混合邊界偵測模型

Abstract

本論文提出一個由生物觀點所啟發的多階次質感邊界偵測演算法。在演算法的發展階段，我們成功地整合了三個重要的視覺元素：明度、質感、色彩。現今相關研究的盲點在於僅從應用觀點出發，無法對整個議題做出完整的探討。有鑑於此，本論文對於人類視覺系統的基本運作模式乃至系統化的整合過程進行相關的基礎研究，針對質感邊界偵測的議題完成了通盤的探討。此外，在多階次的質感特徵萃取過程中，一些為過去研究所忽略但卻極其重要的議題，例如：偽反應（false response）的產生，權重值的選定等等，我們也做了徹底地討論並提出解決的方案。 人類視覺系統能夠有效率地處理視覺資訊的關鍵在於拮抗式的傳送機制，諸如接收域（receptive field）的組成方式以及對比色（opponent color）的形成。本論文參考視覺系統的編碼方式，並以系統化的方式建構出完整的質感邊界偵測演算法。輸入的彩色影像首先被解構為三組對比色軸，並經由高斯（Gaussian）濾波器以及賈伯（Gabor）濾波器萃取質感的一階特徵及二階特徵，輔以本文所提出之適應性權重值決定法則得到兩者對應邊界之權重值，我們可以結合出多階次的質感邊界。經由大量的測試結果，我們發現均勻質感之間的邊界都可以成功而精確地被標定，而對於較不規則或不均勻的質感圖形，演算法仍會找出一些符合我們人眼感受的特性。除了令人滿意的測試結果，本演算法的處理過程極為簡單且直觀，不需導入過多的假設以及任何的訓練過程。相較現有研究，本論文深具應用潛力。

In this thesis, a hybrid-order texture boundary detection technique inspired from human visual system (HVS) was presented. The proposed algorithm integrates three important visual primitives: luminance, texture, and color into a functional system. At present, the related works were developed for specific applications such that an overall investigation of the texture segregation process would be inaccessible. Therefore, the thesis focuses on relevant fundamental researches on HVS and systematic integration to investigate the task of texture boundary detection thoroughly. Moreover, some critical but ignored issues from the procedure of hybrid-order feature extraction, such as false response, weights selection, etc., were also discussed and solved in this thesis. Transmission with antagonism such as receptive field profile and opponent color is the critical point that HVS can effectively process visual information. This thesis employs the encoding form in HVS with systematic integration to build up a complete algorithm for texture boundary detection. Color images are firstly decomposed into three opponent axes and the 1st- and 2nd- order features are extracted by a Gaussian filter and Gabor filters. With the proposed adaptive weights selecting mechanism, the hybrid-order boundary can be obtained. Among extensive tests, boundaries between uniform textures can be detected successfully and accurately. For textures that are non-uniform or non-regular, the results also reflect some meaningful properties which are consistent to human visual sensation. In addition to the satisfying testing results, processing employed in this algorithm is very simple and intuitive with only few assumptions and no training procedure involved. Compared with the present researches, the proposed algorithm has a good application potential.