基于视觉显著性和非监督学习的目标检测

发布时间：2018-06-05 17:48

本文选题：视觉显著性 + 目标检测　；参考：《吉林大学》2017年硕士论文

【摘要】：人类视觉神经系统可以主观地处理看到的自然场景,因为计算能力的限制和高时效性的需求,处理过程会被简化压缩:视神经会选择场景中吸引其注意的部分目标率先进行理解,这种机制被称作选择性注意机制。反过来从能展示自然场景的图像出发,反映场景的图像吸引视神经的性质,被称为视觉显著性。因此,对人类视觉神经系统进行模拟,能够构造出图像的显著性目标检测模型。现在,国内外很多研究团体都在致力于这一工作。本研究针对的图像类型是自然场景类的图像,从颜色空间对比度信息的角度深入探索,主要完成了以下工作:针对已有的基于色彩空间对比度信息的显著性检测方法进行了优化。已知减少颜色数量能够极大加快算法运行效率,指出已有方法在减少颜色数量上存在一定的瑕疵:简单的把RGB三个通道各自进行了量化限制,没有从图像整体的角度针对颜色统计进行量化,这就有可能造成量化后图像一定程度上的失真。明确问题后,提出把八叉树运用到颜色表示中,用以统计颜色出现频率,进而形成调色板完成颜色量化的工作。然后经过直方图加速和去噪平滑操作后得到最终的显著图,因为失真情况已有改善,故可以使显著图较原方法效果更好,而且因为使用了八叉树算法,在时间复杂度上也会有所优化。将本文方法同几种不同原理的经典算法进行了对比,结果证明本文方法效果较好,同时在时间成本上有一定优势。将模糊C均值聚类算法应用于图像处理中,根据聚类结果进行分割。模糊C均值聚类(FCM)算法是一种基于模糊理论的软聚类算法,将该算法应用到图像分割当中,可以获得比传统硬聚类算法更好的效果。对比分析原始图像进行基于FCM算法的分割中选择各个参数的不同情况,提出将显著性检测结果与基于FCM算法的分割过程相结合,通过结果证明,的确可以得到更好地效果。本文提出的方法可以胜任通用显著性目标检测工作,且在效果和时间方面有较为出色的表现;显著性检测与FCM图像分割算法的结合应用提升了两者的效果;本文的工作可以作为图像内容感知和图像检索等应用的基础。
[Abstract]:The human visual nervous system can handle the natural scene subjectively, because of the limitation of computational power and the need for high timeliness, the process will be simplified: the optic nerve will choose the part of the scene that attracts attention in the scene to take the lead in understanding. This mechanism is called the selective attention mechanism. In turn, the natural field can be displayed. The image of the scene, which reflects the nature of the optic nerve, is called the visual significance. Therefore, it simulates the human visual nervous system and can construct a significant target detection model of the image. Now, many research groups at home and abroad are working on this work. The image type is a natural scene class. From the point of view of color space contrast information, the following work has been completed: the significant detection method based on the color space contrast information has been optimized. It is known that reducing the number of colors can greatly speed up the operation efficiency of the algorithm, and points out that the existing methods have some defects in reducing the number of colors. Fault: the quantitative restriction of the three channels of the RGB is simply carried out, and the color statistics are not quantified from the angle of the whole image. This may cause a certain degree of distortion of the quantized image. After the problem, the octree is applied to the color representation to count the frequency of color, and then the color palette is completed. The work of color quantization. Then after the histogram acceleration and de-noising smooth operation, the final significant graph is obtained, because the distortion has been improved, so the significant image can be better than the original method, and because of the octree algorithm, the time complexity will also be optimized. This method is the same with several different principles of the classic. The algorithm has been compared. The results show that the method has good effect and has some advantages in time cost. Fuzzy C means clustering algorithm is applied to image processing and segmentation according to clustering results. Fuzzy C mean clustering (FCM) algorithm is a soft clustering algorithm based on fuzzy theory, and the algorithm is applied to image segmentation. We can obtain better results than traditional hard clustering algorithms. Compare and analyze the different conditions of the selection of the original image based on FCM algorithm, the results are combined with the segmentation process based on the FCM algorithm, and the results prove that the results can be better. The method proposed in this paper can win. The work of universal significant target detection has been performed well in effect and time; the combination of significant detection and FCM image segmentation improves the effect of both. The work of this paper can be used as the basis for the application of image content perception and image retrieval.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】