针对目前大多数显著性检测方法中采用背景种子以及局部区域对比度显著性检测模型的缺点, 本文提出了一种综合考虑局部特征以及全局特征的显著性检测算法。在对图像进行分割之后, 算法首先融合了采用多特征方式生成的背景显著图与采用前景区域对比度方式生成的前景显著图, 之后使用高斯滤波器对融合后的结果进行优化形成局部特征显著图。其次, 在局部特征显著图的基础上提取多种特征的样本集合进行训练, 从而得到全局特征显著图。算法最后将第一步生成的局部特征显著图与全局特征显著图进行结合生成最终的显著图。实验部分验证了算法各部分的有效性, 并且在3个公开数据集上对文章方法与近年来优秀的显著性检测算法进行了对比, 实验结果显示, 本文算法在CSSD数据集上的准确率、召回率以及F-measure分别达到了0.837 5、0.743 4和0.813 7, 在其它数据集上也有良好表现。实验表明, 本文算法能够有效抑制背景区域, 并且高亮前景区域, 更好地检测出显著目标。

Due to the most of existing salient detection methods have some disadvantages on using background seeds and local area contrast salient detection model, a visual saliency detection algorithm named salient detection, which combines local feature and global feature, was proposed. After image segmentation, the algorithm firstly applied a background image created by multi-feature methods and a foreground saliency image created by foreground area contrast method, then, the fusion results was optimized by using Gaussian filter and the local feature saliency image was formed. Secondly, the sample set of various features was collected based on the local feature saliency image for practice and finally the global feature saliency image was obtained. At last, it combined the local feature saliency image produced in the first step with the global feature saliency image and created the final saliency image. In part of experiment, the proposed algorithm showed great results of precision, recall rate and F-measure on CSSD data set, with values of 0.837 5, 0.743 4 and 0.813 7 respectively, the performance on other data set was also perfect. The results show that the proposed algorithm can effectively suppress the background area, highlight foreground area and detect the salient target better.