基于SVM和背景模型的显著性目标检测算法

Salient object detection algorithm based on SVM and background model

针对显著性目标多样性和不确定性,机器学习算法无法检测没有先验信息的图像问题,提出了一种基于图像边缘信息构建背景模型结合SVM分类算法的显著性目标检测算法。该方法对输入图像进行超像素预处理,使像素级转化为超像素级,既抑制噪声,又提高了计算效率。利用图像边缘超像素构建图像的初始背景模型,得到初始显著图。基于SVM算法建立目标和背景的分类模型,结合信息熵评价特征图,迭代优化背景模型,同时得到显著性目标。在公开数据库中进行了测试,实验结果表明,提出的检测算法能够在没有任何图像先验信息的情况下有效地检测出图像中的显著性目标,与流行的5种算法相比,该方法对检测目标的尺度和数量都具有较好的鲁棒性。

In view of the diversity and uncertainty of salient objects,machine learning algorithm can’t detect the image without prior information. A salient object detection algorithm is proposed based on the information of area around image to construct background model and SVM classification algorithm.The input image is preprocessed by superpixel segmentation algorithm,so that the pixel level is transformed into superpixel level,which not only suppresses the noise,but also improves the computational efficiency.The input image is preprocessed by superpixel segmentation algorithm,so that the pixel level is transformed into superpixel level,which not only suppresses the noise,but also improves the computational efficiency. The initial background model is constructed by using the superpixels around image edge,and the initial saliency map is obtained.The classification model of object and background is established based on SVM algorithm,and the background model is iteratively optimized by using the information entropy to evaluate feature map. And the same time,the saliency object is obtained.The proposed algorithm in a public database is tested. Experimental results have been performed to demonstrate that the algorithm can effectively detect salient objects without any prior information.Compared with the five popular algorithms,the proposed method has better robustness to the scale and number of detected objects.