物理学报
物理學報
물이학보
2013年
4期
69-78
,共10页
宋长新*%马克%秦川%肖鹏
宋長新*%馬剋%秦川%肖鵬
송장신*%마극%진천%초붕
图像分割%稀疏编码%聚类%空间约束
圖像分割%稀疏編碼%聚類%空間約束
도상분할%희소편마%취류%공간약속
image segmentation%sparse coding%clustering%spatial constraints
提出了结合稀疏编码和空间约束的红外图像聚类分割新算法,在稀疏编码的基础上融合聚类算法,扩展了传统的基于K-means聚类的图像分割方法.结合稀疏编码的聚类分割算法能有效融合图像的局部信息,便于利用像素之间的内在相关性,但是对于分割会出现过分割和像素难以归类的问题.为此,在字典的学习过程中,将原子的聚类算法引入其中,有助于缩减字典中原子所属类别的数目,防止出现过分割;考虑到像素及其邻域像素具有类别属性一致性的特点,引入了空间类别属性约束信息,并给出了一种交替优化算法.联合学习字典、稀疏系数、聚类中心和隶属度,将稀疏编码系数同原子对聚类中心的隶属程度相结合,构造像素归属度来判断像素所属的类别.实验结果表明,该方法能够有效提高红外图像重要区域的分割效果,具有较好的鲁棒性.
提齣瞭結閤稀疏編碼和空間約束的紅外圖像聚類分割新算法,在稀疏編碼的基礎上融閤聚類算法,擴展瞭傳統的基于K-means聚類的圖像分割方法.結閤稀疏編碼的聚類分割算法能有效融閤圖像的跼部信息,便于利用像素之間的內在相關性,但是對于分割會齣現過分割和像素難以歸類的問題.為此,在字典的學習過程中,將原子的聚類算法引入其中,有助于縮減字典中原子所屬類彆的數目,防止齣現過分割;攷慮到像素及其鄰域像素具有類彆屬性一緻性的特點,引入瞭空間類彆屬性約束信息,併給齣瞭一種交替優化算法.聯閤學習字典、稀疏繫數、聚類中心和隸屬度,將稀疏編碼繫數同原子對聚類中心的隸屬程度相結閤,構造像素歸屬度來判斷像素所屬的類彆.實驗結果錶明,該方法能夠有效提高紅外圖像重要區域的分割效果,具有較好的魯棒性.
제출료결합희소편마화공간약속적홍외도상취류분할신산법,재희소편마적기출상융합취류산법,확전료전통적기우K-means취류적도상분할방법.결합희소편마적취류분할산법능유효융합도상적국부신식,편우이용상소지간적내재상관성,단시대우분할회출현과분할화상소난이귀류적문제.위차,재자전적학습과정중,장원자적취류산법인입기중,유조우축감자전중원자소속유별적수목,방지출현과분할;고필도상소급기린역상소구유유별속성일치성적특점,인입료공간유별속성약속신식,병급출료일충교체우화산법.연합학습자전、희소계수、취류중심화대속도,장희소편마계수동원자대취류중심적대속정도상결합,구조상소귀속도래판단상소소속적유별.실험결과표명,해방법능구유효제고홍외도상중요구역적분할효과,구유교호적로봉성.
A new algorithm for infrared image segmentation is proposed based on clustering combined with sparse coding and spatial constraints. The clustering algorithm is fused on the basis of sparse coding. The traditional image segmentation method based on K-means clustering is extended. The clustering algorithm combined with sparse coding can fuse the local information of image. The inner relationships between pixels are used. However, the problem of over-segmentation and difficulty in pixels classification for segmentation arise. The clustering method is introduced for atoms into dictionary learning. The class number of atoms in dictionary is reduced in order to avoid over-segmentation. The spatial class property information is also introduced by considering the property of the pixel, and the pixels in the neighbor region should have class coherent constraints. An alternate optimization algorithm is proposed to learn the dictionary, sparse coefficients, cluster center and degrees of membership jointly. Then the classes of pixels are estimated by constructing pixel ownership degrees, combining the sparse coefficients and the degrees of membership with the atoms to cluster center. The experimental results show that the important area can be separated well, and the proposed method has good robustness.