摘要： 现有的显著性检测算法多基于图像底层特征,在内容复杂的图像上应用时容易受到干扰。为此,提出一种加入类别先验信息的显著性检测算法。基于深度神经网络生成的特征图谱,选择对预训练集分类结果有正向贡献的部分加权重组,根据保留的空间信息生成显著性图像,结合颜色对比和图像过分割算法进行显著性目标分割。在网络图像组成的公开测试集上的实验结果表明,与IT,SR等算法相比,该算法的平均正确率、召回率和F值都有明显提高。

关键词: 显著性检测, 深度学习, 深度神经网络, 类别先验, 图像特征, 图像分割

Abstract: Most existing saliency detection algorithms are based on low-level image features,therefore the results can be disturbed when applying on complicated images.In order to solve this problem,this paper proposes a saliency detection algorithm using category prior and deep neural network.Based on the feature maps generated by Deep Neural Network(DNN),it selects those parts which contribute to the classification results of pre-training,and generates the saliency map according to the preserved spatial information.With the color contrasting and image segmentation,the test images get the saliency segmentation.Experimental results on public dataset of Internet images show that compared with IT and SR algorithms,the average exact rate,call rate and F value of this algorithm have obvious improvement.

Key words: saliency detection, deep learning, Deep Neural Network(DNN), category priori, image features, image segmentation

中图分类号: 

• TP391.41