深度网络的显著性目标检测

Salient object detection based on deep network

显著性目标检测作为图像的预处理模块,可用于目标检测,目标跟踪及目标分割等各种方面。针对目前显著性目标检测算法精度不高,目标边缘模糊等问题,本文提出深度网络的显著性目标检测算法。算法包括两个互补部分,多尺度全卷积神经网络与背景先验,全卷积神经网络能生成像素级的显著图,解决图像边缘模糊的问题,而背景先验可有效抑制背景冗余信息,最后利用条件随机场将两者进行最优权重的结合。实验表明,通过定性和定量的比较,本文算法的准确度和精确度均得到了一定的提高,且有效解决了目标边界模糊的问题。

Salient object detection is used as an image preprocessing module for object detection,object tracking and object segmentation.For the current salient object detection algorithm is not accurate and the object edge is fuzzy,we propose a salient object detection algorithm based on deep network.The algorithm consists of two complementary parts,the pixel-level convolutional neural network and the background prior,which were combined with the optimal weights of the conditional random field.Experiments show that the accuracy and accuracy of the proposed algorithm are improved by qualitative and quantitative comparison,and the problem of target boundary blur is effectively solved.