霍夫空間轉換應用與服裝規律重複性飾紋之偵測

Abstract

日常生活中，人們所穿著的服裝中經常出現各式各樣的重複性特徵，例如橫條紋、直條紋、格紋以及千鳥格紋等等重複性圖案，不論是在街道上或是在伸展台走秀中均可以看到這些影像。 本論文以前述之觀察為基礎，再利用霍夫空間轉換偵測影像中之直線的特性加以應用，作一自動化偵測影像中人物服裝上之重複性特徵系統，試擷取出影像中可能為服裝上重複性特徵之區塊。 本系統運作略分為以下幾個階段： (1) 以特定之水平、垂直、45度以及135度方向作邊緣偵測，初步定出四個角度的分布圖。 (2) 四個角度的邊緣偵測影像經過門檻篩選後，將影像經過計算轉換到霍夫空間中。 (3) 在霍夫空間中，四個角度的影像分別針對各自角度作特定的門檻篩選機制，把篩選過後的資料存起來。 (4) 在四個霍夫空間中經過篩選過後所存之資料，經過計算畫出相對應之四組具特定方向的線段影像。 (5) 在四組線段中分別用特定機制加上權重，並且加以分群。四個線段影像經過分群篩選後分出四個主要群組。 (6) 最後在四個主要群組中作整合的機制，完成重複性特徵區塊之標定。

經由測試觀察，部分街道中人物服裝以及伸展台走秀人物服裝之重複性特徵可由本系統擷取出來，驗證出霍夫空間轉換工具不僅僅偵測影像中的直線段，其亦可作重複性特徵之偵測的相關應用。

Perception of repetitive imagery patterns is a common experience to all in daily life, the detection of which is useful to many image-oriented applications, and as a result is chosen as the subject of this thesis. It is noted that, in an edge-featured map derived from an image with contents of repetitive patterns, the fragments of edges associated with repetitive-patterned areas appear to lie on the same “line”, which seems to be a general phenomenon through the examination on images in large amounts. Based on such observations, a pilot system exploiting Hough transform for identifying those fragments of edges through which a common line may pass is established. The system operates in 5 steps as follows: [1] Four edge-featured maps are derived from the input, each being with edges in specific orientation: 0゜, 90゜, 45゜, and 135゜. [2] The four maps are transformed into Hough-space, where thresh-holding is taken to select well-responded areas. The transformation is devised in such a way that every image entity in Hough-space is associated with (x, y) coordinates of its counterpart in the image-space. [3] Inverse transform the 4 Hough-space maps back to 4 maps in image-space, in which all in-line fragments are respectively recovered. [4] Clustering of in-line fragments in each of the 4 image-space maps is conducted, where a weighting process is devised so that regions of potential targets may be discerned from uninterested ones via k-mean clustering. [5] A final fusing of the 4 maps of in-lined oriented-fragments, where process of congregating mutual supportive/antagonistic information among 4 maps is performed, concludes the entire process.

Initial experiment on a set of images shows encouraging results from the system when extracting repetitive-patterned dresses/fashion where in-lined fragments are relatively close to one another in testing images. As for images of buildings, it is found that refinement of weighting process in [4] is needed for the cases where in-lined fragments appear to be not so tightly located, as shown in such regular architectural structures as window frames and column-and-beam, etc..