對於果蠅腦嗅覺小球的二維影像比對及半自動邊緣偵測

Glomeruli of the Drosophila Brain：2D Image Analogy and Semi-automatic Edge Detection

郭主引(Ju-Yin Guo)

指導教授 ： 陳永昌

國立清華大學/電機資訊學院/電機工程學系/碩士(2010年)

摘要

對於果蠅的大腦研究一直是個顯著的議題，其中以牠觸角內腦葉的嗅學小球功能最備受矚目。科學家想藉著建立標準的嗅覺小球模型以致力收集數據；但是在立體空間中要對不同果蠅的嗅覺小球做切割，著實是一件累人又複雜的工作。如果平面上的邊緣能半自動地被找出來、並在立體空間中堆疊起來，當作可靠的導引，一個初始模型就可以更準確地朝它們變形，而且能減低人為的負擔。

為了達成半自動偵測嗅覺小球影像邊緣的需求，我們從影像比對的概念獲得靈感，設計一種方法紀錄於此論文中。使用者可提供一組參考影像當作事前資訊，包括一張原始影像以及對應它、畫好的邊緣圖。然後一張新進影像的區塊會和一個資料庫裡的多個區塊做比對；這個資料庫是藉著剪下參考影像不同方向的區塊建立的。我們採用兩種特徵：一種是對影像做高斯函式的偏微分、另一種是經過非最大值抑制的邊緣影像。甚且，為了計算兩種特徵的權重，我們把該比對過程公式化成迴歸分析的問題、再用解線性系統的方法求出權重解。另外，為了改善比對差異產生的盲點，我們設計一個門檻來預先保留較重要的邊緣。最後，一個僅需一組參考影像和一個調整門檻的架構於此產生；能夠進行影像比對、半自動地偵測邊緣，並將它們一步步地合成出來。

關鍵字

邊緣

並列摘要

Cerebral researches of Drosophila melanogaster have been prominent issues. Among them the functionalities of the glomeruli in antenna lobe of D. melanogaster draw the greatest attention. Scientists are striving to gather statistics through a constructed standard glomerular model. However, it is indeed a laborious and complicated process to perform 3D segmentation of glomeruli from various flies. If edges from 2D space are semi-automatically given and stacked up as reliable guidance in the 3D space, a source model can warp to them and be segmented with higher accuracy and less manual effort.

To achieve the desire of semi-automatic edge detection on glomerular images, we developed a method inspired from the concept of ‘image analogy’ in this thesis. Users could provide a source pair, including one original image and the other the corresponding edge map, as the prior guidance. Patch-wise analogy was then applied between patches from the input image and ones from the database, which was constructed by cutting the source pair to patches with different orientations. Two kinds of feature related to gradient were adopted: one was Gaussian partial derivative, and the other was the non-maximal-suppressed edge map. Furthermore, on purpose of calculating the best weights of each feature, we formulated the analogizing process to a regression problem and solved the weights by least squares approach. We also set a threshold to keep the important edges in advance as well as improve the blind spot of the analogized error. Finally, a framework that only asked for single source pair and an adjusted threshold was formed. It performed image analogy and edge detection semi-automatically, and synthesize reliable edge maps step by step.

並列關鍵字

edge

參考文獻

[17] Christopher M.Bishop, “Pattern Recognition and Machine Learning”, p.137~p.147.

[14] Sameer Agarwal and Serge Belongie, “Segmentation by Example”, 2002.

[1] Gregory SXE Jefferis, Elizabeth C Marin, Ryan J Watts, and Liqun Luo, “Development of neuronal connectivity in Drosophila antennal lobes and mushroom bodies”, Current Opinion in Neurobiology, 12, 80-86, 2002.

[2] Africa Couto, Mattias Alenius, and Barry J. Dickson, “Molecular, Anatomical, and Functional Organization of the Drosopjila Olfactory System”, Current Biology, vol.15, 1535-1547, 2005.

[3] P.P.Laissue, C.Reiter, P.R.Hiesinger, S.Halter, K.F.Fischbach, and R.F.Stocker, “Three-Dimensional Reconstruction of the Antennal Lobe in Drosophila melanogaster”, The Journal of Comparative Neurology, 405, 543-552, 1999.