高技术通讯
高技術通訊
고기술통신
HIGH TECHNOLOGY LETTERS
2014年
10期
1-11
,共11页
苗军%卿来云%陈熙霖%马志国
苗軍%卿來雲%陳熙霖%馬誌國
묘군%경래운%진희림%마지국
超像素分割%归一化割%局部图划分%传导率%局部懒惰随机游走(LLRW)
超像素分割%歸一化割%跼部圖劃分%傳導率%跼部懶惰隨機遊走(LLRW)
초상소분할%귀일화할%국부도화분%전도솔%국부라타수궤유주(LLRW)
superpixel segmentation%normalized cut%local graph partition%conductance%local lazy random walk (LLRW)
针对基于图论的超像素分割方法缺乏超像素紧凑性控制和运算复杂度过高的问题，提出了一种基于局部懒惰随机游走（LLRW）的超像素分割方法，并将超像素分割形式化为像素邻接图的局部划分问题，提出了一种直观的分割质量度量。该方法首先将均匀平铺的六边形重心作为超像素种子点初始位置；然后利用局部随机游走算法计算种子点与周围像素的相关程度，将其最相关种子点的标号赋予该像素；最后计算新的超像素重心，并将其作为下一轮迭代的种子点位置，通过若干次迭代逐步优化超像素分割结果。此算法具有线性的时间复杂度和线性的空间复杂度，同时超像素分割质量具有理论保证。通过标准数据集上的实验证明，该方法不仅能够较好地保持图像边界，还可以保证超像素的紧凑性，从而达到理想的超像素分割效果。
針對基于圖論的超像素分割方法缺乏超像素緊湊性控製和運算複雜度過高的問題，提齣瞭一種基于跼部懶惰隨機遊走（LLRW）的超像素分割方法，併將超像素分割形式化為像素鄰接圖的跼部劃分問題，提齣瞭一種直觀的分割質量度量。該方法首先將均勻平鋪的六邊形重心作為超像素種子點初始位置；然後利用跼部隨機遊走算法計算種子點與週圍像素的相關程度，將其最相關種子點的標號賦予該像素；最後計算新的超像素重心，併將其作為下一輪迭代的種子點位置，通過若榦次迭代逐步優化超像素分割結果。此算法具有線性的時間複雜度和線性的空間複雜度，同時超像素分割質量具有理論保證。通過標準數據集上的實驗證明，該方法不僅能夠較好地保持圖像邊界，還可以保證超像素的緊湊性，從而達到理想的超像素分割效果。
침대기우도론적초상소분할방법결핍초상소긴주성공제화운산복잡도과고적문제，제출료일충기우국부라타수궤유주（LLRW）적초상소분할방법，병장초상소분할형식화위상소린접도적국부화분문제，제출료일충직관적분할질량도량。해방법수선장균균평포적륙변형중심작위초상소충자점초시위치；연후이용국부수궤유주산법계산충자점여주위상소적상관정도，장기최상관충자점적표호부여해상소；최후계산신적초상소중심，병장기작위하일륜질대적충자점위치，통과약간차질대축보우화초상소분할결과。차산법구유선성적시간복잡도화선성적공간복잡도，동시초상소분할질량구유이론보증。통과표준수거집상적실험증명，해방법불부능구교호지보지도상변계，환가이보증초상소적긴주성，종이체도이상적초상소분할효과。
A novel approach to generate superpixels based on local lazy random walk (LLRW) was proposed to improve the compactness of superpixels and reduce the computational complexity. The superpixel segmentation was formulated as a problem of local partition of pixel adjacency graphs, and an intuitive quality measure for superpixel segmentation was given. The LLRW approach firstly initializes the centroids of uniformly titled hexagons as the positions of superpixel seeds, and then uses the local lazy random walk (LLRW) algorithm to calculate the correlation between nearby pixels and superpixel seeds, and sets the label of each pixel as the label of its most correlated seed. Finally, it calculates the new centroids of superpixels as the next iteration’s seed positions, and iterates these steps to refine the superpixel segmentation result. This algorithm has the linear time complexity and the space complexity, as well as a theoretical guarantee on the quality of superpixels. The experimental results show that the new method can preserve smooth boundaries and generate compact superpixels, so it is an ideal algorithm for real world applications.
