尋找自然景觀中之野生動物

Abstract

在本文中我們利用若干已知之人類視覺系統特性，建構一套在自然景觀中偵測野生動物的系統。這套系統在偵測自然景觀中的野生動物時，相較其他應用電腦技術但缺乏任何人類視覺系統相關知識支援的系統來說，對影像條件的要求限制較少。對人來說，我們可以在自然景觀中輕易找出適當大小的野生動物。但對機器來說，這並不是一件容易的事。 從以往對人類視覺系統所作的研究中，我們利用眼睛視網膜上的神經結對光線刺激的反應呈現一種 center-on 或 center-off的現象來作為系統對影像處理之基礎。所謂 center-on，就是用來描述某些神經結只在光線刺激其中心區域時有反應，而 center-off，則是描述某些神經結只在光線刺激其周圍區域時有反應。我們利用一些程式模組來模擬人類視覺系統中 center-on 與 center-off 的神經結，抽取出影像中相符合的視覺訊息，再利用我們所設計出來的運算元來處理影像中灰階與對比的變化，經由一個 bipolarized convolution 的程序，將抽取出來的影像資料做疊合的處理，藉此來找出影像中的物體。 根據這些簡單的運作，我們的系統可以用很低的計算成本找出野生動物所在的位置。更重要的是，這套系統沒有使用任何與影像主體有關的模型，只是用一種類似人們觀看物體的直覺反應方式來找出影像主體，並能將影像主體之外形以和吾人視覺經驗相符的方式予以捕捉並顯現出來。最後我們拿這套系統所產生的結果與用傳統影像切割方法所得到的結果做比較。

An automated system for detecting the wildlives in generic scenes is proposed in this thesis based on certain established knowledge about the behaviors of human visual system. The system can detect the wildlives in generic scenes with less requirements on the input restrictions, as compared to most extant systems which employ computing techniques of all sorts but lacking supports from the biological grounds. It is easy for people to see the wildlives of proper scales in generic scenes, but it is definitely not easy for machines to do the same job. According to the research of human visual system, it’s known that, in the retinas of human eyes, center-on ganglion cells become active when the light stimulates their center regions and remain silent when the light stimulates their surround areas. And center-off ganglions cell will be active and silent in opposite conditions. Some functional modules emulating the behaviors of center-on and center-off ganglion cells are employed at early stages of system process to make the objects stand out from the background. These operations are done by primitive operators working in either contrast or gradient space and are embedded in a process called bipolarized convolution for contrast feature extraction and for subsequent fusion in the detecting process. With the aid of findings in contrast perception and focus of visual attention of human visual system, though very limited ones, we demonstrate that the system is able to locate the wildlives of proper scales in the images accordingly. Most importantly, the system uses no object models of any kinds and detects objects in a holistic style, very similar to what a human viewer see. Finally we compare the results generated by our system with those of another system using conventional region-split-and-merge technique for image segmentation incorporated with a set of human visual attention behaviors.