基于深度神经网络的图像显著性预测方法研究与实现

发布时间：2018-03-05 01:23

本文选题：显著性预测　切入点：视觉注意机制　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：人类在长期的进化中具有了高度发达的视觉注意机制,拥有在复杂的环境中快速选择性的发现感兴趣目标的能力。模拟这种视觉注意机制引起了在心理学、神经系统,尤其是计算机视觉领域里的许多研究者们的重视。图像数据信息是视觉信息的主要载体,当注意一张图片的时候,人类的视觉注意机制使得他们在图片上的注视点会落在更能刺激视觉的特殊区域,这些特殊区域就是图像中比较显著的区域。显著性预测就是通过模仿人类视觉注意机制来提取图像的重要信息,并应用于新的人工智能领域。本文使用深度学习技术建立了图像显著性预测的两种模型。在构建深度学习模型时,我们综合了深度卷积神经网络的优点和传统的显著性原理。我们在训练深度神经网络时,综合考虑了图像的High-Level与Low-Level语义信息,提出了多尺度的卷积神经网络;综合传统显著性检测中基于全局对比和局部对比的方法,提出了循环的卷积神经网络,从多种层次上考虑了图像所蕴含的信息。本文使用的两种网络模型,多尺度的卷积神经网络和循环卷积神经网络,对图像处理与识别的其它领域也有着借鉴意义。本文提出的周期性训练方法通过融合原图像和初步训练的显著图,可以达到自动优化显著图的目的。除此之外,本文通过模糊化显著图的办法,使得两种模型的实验结果更加精确。本文以Caffe深度学习框架为平台实现两种深度神经网络模型,使用Matlab,Python等软件预处理图像、导入和导出模型结果。我们对使用的两种模型在MIT,SALICON与iSUN三种数据库上进行了实验,并从定性和定量的角度进行了评估。本文提出的两种模型最终的实验结果均优于传统模型和普通的卷积神经网络模型所预测的显著性结果。
[Abstract]:Humans have evolved over a long period of time with highly developed visual attention mechanisms, with the ability to quickly and selectively detect objects of interest in complex environments. Especially the attention of many researchers in the field of computer vision. Image data information is the main carrier of visual information, when paying attention to a picture, Human visual attention mechanisms allow their gaze on the image to fall in a particular area that is more visually stimulating. These special regions are the more significant areas in the image. Salience prediction is to extract the important information of the image by imitating the human visual attention mechanism. In this paper, two models of image salience prediction are established by using depth learning technology. We synthesize the advantages of deep convolution neural network and the traditional significance principle. When we train depth neural network, we consider the semantic information of image High-Level and Low-Level, and propose a multi-scale convolution neural network. Based on the methods of global contrast and local contrast in traditional salience detection, a cyclic convolution neural network is proposed, which considers the information contained in the image from various levels. Two network models are used in this paper. The multi-scale convolution neural network and cyclic convolution neural network can also be used for reference in other fields of image processing and recognition. In addition, this paper makes the experimental results of the two models more accurate by the method of fuzzy salience map. In this paper, two depth neural network models are implemented on the platform of Caffe depth learning framework. Using Matlab Python and other software to preprocess image, import and export model results, we have done experiments on three kinds of database, such as MIT SALICON and iSUN. The experimental results of the two models presented in this paper are better than those predicted by the traditional model and the conventional convolution neural network model.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】