【作者】 丁玲；

【导师】 李建华；

【作者基本信息】 大连理工大学 ， 电子与通信工程（专业学位）， 2017， 硕士

【摘要】 随着互联网技术的飞速发展,图像采集设备如数码相机和智能手机等的广泛普及,每天都会有海量的图像数据产生,在这些数据中掺杂着大量冗余信息。人们在观测这些图像数据时会首先注意图像中感兴趣的区域,进而对其进行优先处理。显著性检测就是利用计算机技术模仿人类的视觉注意机制,自动提取图像中显著区域,这不仅可以大大压缩计算机待处理的数据,节约计算机的存储空间,同时可以提高计算效率。目前显著性检测方法已获得许多成果,但更多仅适用于单图检测,不适用于多图协同显著性检测。多图协同显著性目标检测一般应用于从多幅图像或多个视频中,寻找相同或相似的显著性目标,其强调显著性和协同性。本文在研究已有工作基础上,提出了两种协同显著性目标检测算法:基于多聚类的协同显著性目标检测算法、基于多特征融合多示例学习的显著性目标检测算法。观察发现,当图像场景简单时,选用较少的聚类数目就可以有较准确的聚类结果;而当图像场景复杂时,选用较多的聚类数目才会有准确聚类结果,为了能对不同场景实现自主选择更准确的聚类数目,本文提出了基于多聚类的协同显著性目标检测算法。该算法将多聚类数目下生成的协同显著性结果进行不同权重的融合,即对每一场景图像执行多个聚类数目下的聚类,再在每一聚类数目下计算该聚类数目对应的置信度,置信度越高,则在最后融合过程中赋予越大的权值。另外,本文提出了基于多特征融合多示例学习的显著性目标检测算法。该算法考虑到不同特征检测出的结果对于不同场景图像的贡献度不同,因此构建了如下算法步骤:首先提取示例构建正负包,提取所有包的颜色、纹理、对比度、VGG16特征,分别进行多示例学习,测试出的多个结果由低秩分解重构误差计算出的权重进行融合,再分别经图内的协同显著值排序扩散和组内的类间协同显著值传播,生成最终的结果图。两种算法皆采用了初始显著图,来源于现有的显著性检测算法。实验部分,本文将这两种算法应用于现有显著性检测算法的后处理,并对性能的提升与否做了定性和定量的评判,同时在两个公开的协同显著性检测数据库上与其他算法进行了对比。更多还原

【Abstract】 With many image devices such as digital cameras and smart phones widely promoted,massive image data naturally generates,which is including a lot of redundant information.Saliency detection can extract areas of images which people are interested in,and the extracted areas will be processed by the computer priority.On the one hand saliency detection makes the data to be processed by computer greatly compressed and saves the computer’s storage space;on the other hand it improves the computing efficiency.The saliency detection method has obtained many results,but it only applies to single image detection at present and does not apply to multi-images co-saliency detection.Co-saliency detection is generally used to find the same or similar significance areas from multiple images or multiple videos,which emphasizes saliency and high repetition rate.There are few studies in this region at present.This paper presents two co-saliency detection algorithms,which are described below.The first algorithm is based on multi-clustering fusion.We observe that the clustering accuracy is different for different image scenarios under the same number of clusters.When the scene in the picture is simple,less clustering number will be accurate.Otherwise,the complex scene will be detected accurately under larger clustering number.To choose the better clustering result automatically for different scenes,we present to design weights for different clustering weak co-saliency results.That is,first we perform multi-clustering by set different numbers for clusters and get the clustering accuracy by calculating the confidence in every number.Confidence value is higher,the corresponding weight is bigger.Second,after we get the weak maps under different clustering numbers,we use the confidence values to fuse the weak maps to get the strong co-saliency map.We get the final map by fusing the initial saliency map and strong co-saliency map.The second algorithm is based on multi-features fusion multiple instance learning.In this algorithm we propose that the co-saliency results detected by different features make different contributions to distinct scenes,which mainly depends on the similarity between foreground and background.The specific steps of the algorithm are as follows: first,extract the positive and negative bags.Then extract color,texture,contrast,VGG16 features of bags to train and test KI-SVM network.Multi-results are fused by the weights to generate strong co-saliency maps,and weights are calculated by low-rank recovery errors.After manifold ranking in image and propagation co-saliency value in group,we get the final co-saliency maps.The two algorithms all use the initial saliency map,which is based on the existing saliency detection algorithm.In the experimental part,the two algorithms are applied to the post-processing of existing saliency detection and we evaluate the performance in qualitative and quantitative parts.At the same time we compare our algorithms with other algorithms in two publicly co-saliency datasets.更多还原