目的 感兴趣区域检测是图像处理领域的关键技术.人类视觉系统处理一个较为复杂的场景时,会首先将其视觉注意力集中于该场景中的几个特定对象上,这些对象被称为感兴趣区域.在图像处理和分析过程中,感兴趣区域检测模拟人类视觉,能够快速、准确抓住图像重点,降低图像处理计算量,有效提高计算机信息处理的效率.因此感兴趣区域检测对于图像分析和理解有着重要意义.为此,提出一种基于低层次图像信息与中层次图像信息相结合的自底向上的感兴趣区域检测方法.方法 首先通过彩色增强Harris算子检测角点进而得到凸包边界,通过凸包区域与超像素聚类结果计算中层次信息粗略显著图;然后将图像从RGB空间转换到CIELab空间,使用差分滤波器对图像进行滤波,得到低层次信息粗略显著图;最后将低层次图像信息与中层次图像信息进行加权融合得到图像的显著图.结果 在微软亚洲研究院提供的公开数据库MSRA上验证了本文方法的有效性,根据该数据库人工标记的真值评价本文方法的检测效果,并与其他方法进行对比.其他方法的显著图是由其作者提供的源代码得到.在主观分析和客观判断两个方面的本文方法可有效抑制背景噪声,检测出的显著物具有均匀显著度,且边缘清晰.结论 本文方法是一种有效的图像预处理方法.

Objective The detection of region of interest (ROI) is the a key technique in image processing. Human visual system focus on a few objects in a complicated natural environment. These objects are called region of interest. The model of region of interest detection can simulate the human visual system and accurately compute the saliency area in image processing. This model can improve the efficiency of computer processing and reduce calculation complexity. Thus, the detection of region of interest is of great significance.Method A bottom-top ROI detection method is proposed based on low level image cues combined with middle level cues. First, the middle level coarse saliency region is obtained via a convex hull of corner detected by boosting Harris and superpixels clustering. The original image is then transformed from RGB color space to CIELab color space, and the difference of Gaussian filter method is presented to obtain the low level coarse saliency map. Eventually, the saliency map of the initial image is obtained by fusing the two coarse saliency maps.Result Extensive experiments on the large data set coming from Microsoft Asian research institution show that our method performs better than state-of-the-art algorithms. For fair evaluation, the results obtained via the five methods are based on the source codes provided by the authors. Both a subjective and objective evaluations of the proposed method compared with the other five methods are presented. The subjective comparison illustrates that our method provides accurate location, well-defined boundaries, uniform highlight, and full resolution saliency map. Moreover, the objective comparison via precision-recall curve shows that our method performs well in precision. Experiments show that this method can clearly highlight the whole salient object via reduced degrees of saliency levels, significantly alleviate the influence of false positive pixels, and obtain well-defined boundaries.Conclusion In conclusion, our method can be generally exploited as an image preprocessing method.