基于全卷积神经网络和多核学习的显著性检测

Saliency detection based on fully convolutional network and multiple kernels learning

针对显著性检测中特征选择的主观片面性和预测过程中特征权重的难以协调性问题,提出了一种基于全卷积神经网络和多核学习的监督学习算法。首先通过MSRA10K图像数据库训练出的全卷积神经网络(FCNN),预测待处理图像的初步显著性区域;然后在多尺度上选择置信度高的前景、背景超像素块作为多核支持向量机(SVM)分类器的学习样本集,选择并提取八种典型特征代表对应样本训练SVM;接着通过多核SVM分类器预测各超像素显著值;最后融合初步显著图和多核学习显著图,改善FCNN网络输出图的不足,得到最终的显著性目标。该方法在SOD和DUT-OMRON数据库上有更高的AUC值和F-measure值,综合性能均优于对比方法,验证了该方法在显著性检测中准确性的提高,为目标识别、机器视觉等应用提供更可靠的预处理结果。

Considering the one-sidedness caused by individual subjectivity in feature selection and the difficulty in coordinating feature weights,this paper proposed a supervised saliency detection algorithm based on fully convolutional neural network(FCNN)and multiple kernels learning(MKL).Firstly,the algorithm trained a FCNN with MSRA10K dataset,and utilized the FCNN to predict coarse saliency regions.Then the algorithm extracted and labeled multi-scale superpixels with their corresponding saliency scores.Multi kernels support vector machine further treated eight typical features extracted from each of those superpixels as the training set and produced a finer saliency map.Finally,the algorithm fused the two saliency maps to generate the final saliency map.The method reaches higher area under the roc curve(AUC)value and F-measure value on SOD dataset and DUT-OMRON dataset,the comprehensive performance is better than any of the compared methods,which indicates the improvement on detection precision and can provide more reliable pre-processing result for object recognition and machine vision.