对区域的边界和物体边缘像素使用聚焦操作来计算区域显著特征,采用全局颜色显著性计算全局显著特征,基于卷积神经网络(CNN)融合区域显著特征和全局显著特征,获得最终的显著图,同时采用循环结构网络,多次参考周围环境信息,剔除噪声特点。在MSRA图像库和ECSSD图像库中测试所提算法,其平均精度和平均召回的调和平均值、平均误差均优于当前流行算法。

The prominent features of the salient region are determined by focusing on the regional boundary and the object's edge pixels. Further, the uniqueness of the salient global color is used to calculate global features. Finally, the salient region is obtained using the convolutional neural network (CNN) model based on the regional and global salient features. Adopting a circular structure network is critical to eliminate the noise characteristics by referring to the surrounding environment information for multiple times. The proposed algorithm is tested using the image libraries of MSRA and ECSSD and it is found that its harmonic mean and average error associated with the average precision and recall are better than those of the current popular algorithms.