刑侦现勘图像特征提取与语义标注技术研究

发布时间：2019-02-14 16:33

【摘要】：在当前的国内公共安全领域中,基本上仍然采用纯手工的方式对刑侦现勘图像进行标注,这种方式不仅效率低下,并且带有强烈的主观性,针对这一现状,本文在阅读大量国内外参考文献的基础上,围绕图像特征提取和语义标注,提出了一些自己的改进方法,期望能够实现刑侦现勘图像的自动语义标注,减轻公安干警的工作强度。首先深入研究了颜色和纹理两种底层的图像特征,在这两种特征的基础上,通过加入不同的权值系数,提出了融合的特征提取算法。经过大量的实验,将实验结果中表现最好的权值作为融合特征提取算法的参数,以之为基础,借鉴图像分类思想,将图像语义标注问题转化为图像分类问题,通过特征直方图相似度对比的方式实现图像分类,进而实现刑侦现勘图像的语义标注。实验结果表明,融合的特征算法比两种单独的特征算法对于图像语义标注准确率有明显提高。然后对SIFT和SURF特征进行了研究,针对SURF特征在图像的尺度和旋转变化方面表现不够理想的情况,借鉴SIFT特征的思想,以SURF特征为基础,结合高斯金字塔模型,提出了 GP-SURF特征提取算法。核心思想是在尺度空间构造阶段采用图像尺寸改变的高斯差分金字塔,模拟人眼由远及近视物这种图像尺度改变的情况,期望能够克服图像尺度和旋转变化的影响,改善刑侦现勘图像的特征表示。实验结果表明,GP-SURF特征明显提高了刑侦现勘图像的语义标注准确率。接着深入研究了词袋模型,借助这一模型,将提取的图像特征采用一种聚类算法进行聚类,构成视觉单词,视觉单词的集合组成视觉词典,然后通过支持向量机对其进行训练,得出各类图像的分类超平面,接着以两两比较的方式构造了支持向量机分类器,通过这个分类器对图像进行分类和标注,从而实现了一种图像语义标注模型。通过实际的实验过程,验证了该图像语义标注模型的有效性。最后设计实现了一个“案件图像管理与检索系统”,将提出的颜色和纹理融合特征与GP-SURF特征再次融合,得到再融合特征HL-GS,以此特征为基础,通过再融合特征语义标注模型实现了刑侦现勘图像的语义标注,并且将其应用到了“案件图像管理与检索系统”中的图像语义标注模块和图像检索模块。
[Abstract]:In the current field of public security in China, we still use the pure manual method to mark the image of criminal investigation. This method is not only inefficient, but also has strong subjectivity. In view of this situation, On the basis of reading a large number of references at home and abroad, this paper puts forward some improved methods around image feature extraction and semantic annotation, which is expected to realize automatic semantic annotation of forensic survey images. Reduce the intensity of the work of public security officers and policemen. Firstly, two kinds of bottom image features, color and texture, are studied in depth. Based on these two features, a fusion feature extraction algorithm is proposed by adding different weight coefficients. After a large number of experiments, the best weights in the experimental results are taken as the parameters of the fusion feature extraction algorithm. Based on them, the image semantic tagging problem is transformed into the image classification problem based on the image classification idea. Image classification is realized by comparing the similarity of feature histogram, and the semantic annotation of forensic prospecting image is realized. The experimental results show that the fused feature algorithm can improve the image semantic tagging accuracy obviously compared with the two single feature algorithms. Then, the SIFT and SURF features are studied. In view of the fact that the SURF features are not ideal in terms of the scale and rotation of the image, the idea of SIFT features is used for reference, and based on the SURF features, the Gao Si pyramid model is combined. GP-SURF feature extraction algorithm is proposed. The core idea is to use the Gao Si differential pyramid with image size change in the construction stage of scale space to simulate the image scale change of the human eye from distant to nearsightedness, in order to overcome the influence of image scale and rotation change. To improve the feature representation of the present image of criminal investigation. The experimental results show that GP-SURF features can improve the accuracy of semantic tagging of forensic survey images. Then, the lexical bag model is studied in depth. With the help of this model, the extracted image features are clustered by a clustering algorithm to form visual words, the set of visual words constitutes a visual dictionary, and then it is trained by support vector machine (SVM). The hyperplane of all kinds of images is obtained and then a support vector machine classifier is constructed by pairwise comparison. The classifier is used to classify and annotate images and thus a semantic annotation model of image is implemented. The effectiveness of the image semantic annotation model is verified by the actual experimental process. Finally, a case image management and retrieval system is designed and implemented. The proposed color and texture fusion feature and GP-SURF feature are fused again, and the refusion feature HL-GS, is obtained based on this feature. The semantic annotation of criminal investigation image is realized by remerging the feature semantic annotation model, and it is applied to the image semantic annotation module and image retrieval module in case Image Management and Retrieval system.
【学位授予单位】：西安工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】