应用颜色聚类图像块的多舰船显著性检测

Multi-ship saliency detection via patch fusion by color clustering

由于多舰船目标显著性检测过程容易将边界像素作为背景处理,本文提出了应用颜色聚类图像块的多舰船显著性检测方法。该方法首先检测邻域像素是否具有颜色相似性,并将临近的具有相似颜色的像素聚集在一起作为一个图像块。接着,对获得的图像块进行扩展,使图像块包含很多其他图像块的像素以提高图像块内像素间的对比强度;对边缘像素进行背景索引标记,计算图像块中像素的显著性强度,采用阈值分割方法获得目标显著性区域。最后,基于颜色聚类的图像块存在部分重叠的特点,利用权值对存在叠加的显著性图像进行融合,从而获得多舰船目标整幅图像的显著性检测结果。对获得的多舰船目标图像进行了实验测试,并对本文算法结果和当前比较先进的其它显著性检测算法进行了效果对比。结果显示:提出的利用颜色聚类图像块的舰船显著性检测方法的查全率达到78%以上,准确率达到92%以上,综合评价指标Fβ≥0.7;无论考虑单个指标还是整体指标,本文算法均优于其他对比算法。

Because the boundary pixels are easy to be classified as a background in the multi ship target detecting processing, this paper proposes a multi-ship saliency detection method based on patch fusion by color clustering. Firstly, this method detects the color similarity of the pixels in the neighbourhood, and the adjacent pixels with the similar color are gathered as an image patches. Then, the image patches are expanded to make them include some pixels of other patches, so as to enhance the contrast value of the pixels of patches. Then, edge pixels are marked in the background index to calculate the saliency ability of the pixels in image patches and the threshold segmentation method is used to obtain the saliency region of the target. As the image patches have the features of partial overlap, the weight values are used to fuse the saliency images with the partial overlaps, so that the saliency detection results on a whole image for the multi-ship targets are obtained. The experimental tests are carried out for the multi-ship target images, and the results from the proposed algorithm in this paper and the current advanced detection algorithms are compared. The results show that the proposed method based on patch fusion by color clustering has the recall rate more than 78%, theaccurate above 92%, and its comprehensive evaluation index Fβ is more than 0. 7. Both for comparisons of the single index or the entire indexes in this experiments, the algorithm is superior to other methods.