融合全局与局部特征的相似视频片段快速检测技术研究

发布时间：2018-09-19 14:51

【摘要】：本文在深入了解相似关键帧检测与相似视频检测国内外研究现状的基础上,分析两者存在的不足,并在以下几个方面展开深入研究,取得一定进展。首先,改进了关键帧全局颜色直方图,提出光照-尺度金子塔特征。通过构建关键帧亮度、裁剪尺度空间金字塔,提高该特征对光照、尺度裁剪变换的鲁棒性。实验证明,基于全局光照-尺度金字塔特征的相似关键帧检测查全率较基于全局颜色直方图的检测效果更好,当相似距离阈值?(28)2时,算法查全率达到95.2%。其次,针对SIFT特征数据维度高,计算效率低的缺点,提出一种基于稀疏编码的尺度不变特征加速算法(ScSIFT)。以超完备字典将SIFT特征稀疏表示,同时建立二级特征索引结构,提高特征距离计算速度与检索效率。实验证明,ScSIFT算法与SIFT算法匹配结果类似,但算法运行效率较后者提高了52%。最后,基于本文提出的光照-尺度金字塔特征、ScSIFT算法,结合时序分块顺序算法,提出一种融合全局与局部特征的相似视频片段快速检测算法。该算法融合相似视频片段检测中全局特征运算速度快、局部特征计算精度高的优点。实验对比发现,该算法较传统的算法准确率更高,达到78.2%,同时,算法运行效率较高。
[Abstract]:On the basis of deeply understanding the research status of similar key-frame detection and similar video detection at home and abroad, this paper analyzes the shortcomings of the two methods, and makes some progress in the following aspects. Firstly, the key frame global color histogram is improved, and the illumination-scale gold tower feature is proposed. By constructing the brightness of the key frame and clipping the pyramid of the scale space, the robustness of the feature to the illumination and scale clipping transformation is improved. Experiments show that the detection recall rate of similar key frames based on global illumination scale pyramid features is better than that based on global color histogram. When the threshold of similar distance is? (28) 2, the recall rate of the algorithm reaches 95.2%. Secondly, aiming at the disadvantages of high dimension and low computational efficiency of SIFT feature data, a scale-invariant feature acceleration algorithm (ScSIFT).) based on sparse coding is proposed. The SIFT features are represented sparsely by an overcomplete dictionary and a secondary feature index structure is established to improve the computing speed and retrieval efficiency of the feature distance. The experimental results show that the Ssift algorithm is similar to the SIFT algorithm, but the efficiency of the algorithm is 52% higher than that of the latter. Finally, based on the illumination scale pyramid feature scsift algorithm proposed in this paper and the sequential block sequence algorithm, a fast detection algorithm for similar video segments is proposed, which combines global and local features. The algorithm combines the advantages of fast global feature operation and high accuracy of local feature computation in similar video segment detection. The experimental results show that the accuracy of the algorithm is higher than that of the traditional algorithm, reaching 78.2. At the same time, the efficiency of the algorithm is high.
【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】