基于深度学习理论的纹身图像检测研究

发布时间：2018-05-15 01:24

本文选题：深度学习 + 深度置信网络　；参考：《南昌大学》2017年博士论文

【摘要】：随着图像拍摄设备、智能手机和互联网技术的发展,纹身图像的采集、传播变得越来越容易。伴随着突发事件的发展,纹身同其它生物特征一样,成为对罪犯嫌疑人识别的有力证据。如何对纹身图像进行检测和语义解读,并为相关部门和人员提供有力的证据,已引起安全部门的重视。纹身图像具有明显的图案信息的局部性、内容的复杂性、纹理的清晰性、颜色的单一性、图案Logo标志性、大小形状多样性等特点。这些特点使得纹身图像的检测与识别相对比较困难,同时也使得很难用单一的特征对其描述。深度学习通过逐层追叠形成多层的网络结构,这种结构可以从底层到高层逐层提取到图像的高层特征,从而有效对图像进行表示。这使得深度学习逐渐成为学术界、企业界研究的热点,同时也为纹身图像检测提供了一种新的途径。本文的主要工作是紧密围绕纹身图像的特点和现有的深度学习理论,展开深入的研究,提出针对纹身图像检测的若干改进算法。本文的主要工作和贡献包括以下几个方面:1.四种主要深度学习算法在纹身图像检测中的比较研究。通过分析纹身图像的特点,探讨哪种算法比较适用于纹身检测的研究工作。实验结果表明,在纹身检测方面,四种深度学习算法比传统的方法的性能更好,其中深度卷积神经网络和深度置信网络的性能更为突出。2.基于多特征融合的深度置信网络纹身图像检测改进算法(MF-DBN)。纹身图像的内容复杂性、纹理的清晰性、颜色的单一性等诸多特点决定了单一特征很难对其进行准确描述,同时像素又比较高。传统的DBN算法比较适中小尺寸的图像识别任务,针对这些问题,从多特征融合的视角,设计了一个基于纹身图像的多特征融合深度置信网络改进算法(MF-DBN),有效地解决了纹身图像高维性和单一特征描述不足的问题。在NIST纹身数据集上正确率达到96.89%,比NIST公布的最好正确率提高了0.59%。3.基于视觉词包的深度置信网络纹身图像检测改进算法(BOVW-DBN)。该算法不仅解决了输入高维性的问题,还针对纹身图案的局部性、大小不一性等特点,利用SIFT算法,建立中层词包语义模型,有效的实现对纹身图像的检测与识别。在NIST纹身数据集上正确率达到97.57%,比NIST公布的最好正确率提高了1.17%;在Flickr 10K上的正确率也达到79.27%。4.基于空间金字塔的深度置信网络纹身图像检测改进算法(SP-DBN)。该算法利用SPM模型,针对纹身图像的大小及空间分布等问题,解决了纹身图像的空间信息特征提取问题,实现纹身图像检测。实验结果显示,在NIST纹身数据集上正确率达到97.23%,比NIST公布的最好正确率提高了0.93%;在Flickr 10K上的正确率达到80.46%。5.基于三通道融合的卷积神经网络纹身图像检测改进算法(CFT-CNN)。针对全连接层在不同尺度下的特征抽取能力,首先根据纹身图像的检测问题设计了一个简单的单通道T-CNN模型;在单通道T-CNN模型的基础上,又设计出一个三通道连接层的卷积神经网络模型(CFT-CNN),并应用到纹身图像检测的任务中。同时针对纹身图像的特点做了相应的预处理。在NIST数据集上,CFT-CNN的正确率达到97.87%,比NIST公布的最好结果提高了1.57%,在Flickr 10K数据集上的正确率也达到85.61%。6.基于三通道融合的Faster R-CNN纹身图案检测改进算法(CFT Faster RCNN)。该算法针对纹身图像大小、尺度变化大等问题,在Faster R-CNN的基础上,充分考虑到全连接层在不同尺度下的特征提取能力,在ROI池化后增加一个三通道的全连接层,解决了纹身图像位置识别困难的问题。实验结果表明,在NIST数据集上,本算法比Faster R-CNN算法在MAP上提同了3.35%,在IOU上提高了4.59%。最后,对基于DBN和CNN的纹身图像检测改进算法进行了对比研究。结果表明改进算法在NIST上性能有所提升,其中改进的CNN算法的效果更好。在改进的DBN方法中,由于词包模型是在SIFT特征基础上对纹身图像进行表示,在小数据集上更好。SP-DBN算法由于考虑到空间信息,对大一些的数据集的效果相对较好。从NIST和Flickr两个数据集的实验结果来看,NIST数据集存在一定局限性,Flickr数据集更接近现实环境。
[Abstract]:With the development of image photographing equipment, smart phone and Internet technology, the collection of tattoo images has become more and more easy. With the development of unexpected events, the tattoo, like other biological features, has become a powerful evidence for the identification of criminal suspects. How to detect and interpret tattoo images and to be related departments and people The staff provide strong evidence, which has aroused the attention of the security department. The tattoo image has the features of obvious pattern information, the complexity of the content, the clarity of the texture, the singleness of the color, the Logo logo of the pattern, the diversity of the size and shape, which make the detection and recognition of the tattoo image relatively difficult and also make it possible It is difficult to describe it with a single feature. Deep learning forms a multilayer network structure by overlapping layer by layer. This structure can be extracted from the bottom to the high level to the high level feature of the image, thus effectively expressing the image. This makes the deep learning gradually become a hot topic in the academic circle, and also for the tattoo image inspection. The main work of this paper is to focus on the characteristics of the tattoo image and the existing depth learning theory, and to carry out an in-depth study and propose some improved algorithms for the tattoo image detection. The main work and contributions of this paper include the following aspects: 1. the four main depth learning algorithms are in tattoo image detection. By analyzing the characteristics of the tattoo image, this paper discusses which algorithm is suitable for the research of tattoo detection. The experimental results show that, in the aspect of tattoo detection, the performance of the four depth learning algorithms is better than the traditional method, and the performance of the deep convolution neural network and the depth confidence network is more prominent based on the.2. based on the tattoo detection. The improved algorithm for multi feature fusion of deep confidence network tattoo image detection (MF-DBN). The complexity of the content of the tattoo image, the clarity of the texture, the singleness of the color, and so on, determine that the single feature is difficult to accurately describe it, and the pixels are relatively high. The traditional DBN algorithm compares the image recognition task with the small size and size. To these problems, from the perspective of multi feature fusion, a multi feature fusion depth confidence network improvement algorithm (MF-DBN) based on the tattoo image is designed, which effectively solves the problem of the lack of high dimension and single feature description of tattoo images. The accuracy of the NIST tattoo data set is 96.89%, which is 0. higher than the best correct rate published by NIST. 59%.3. based on the visual word packet, the improved algorithm (BOVW-DBN) of the depth confidence network tattoo image detection (BOVW-DBN). This algorithm not only solves the problem of high dimension of input, but also aims at the features of the locality and size of the tattoo pattern, and uses the SIFT algorithm to establish the middle word packet semantic model, effectively realizing the detection and recognition of the tattoo image. In NIST The correct rate of the tattoo data set is 97.57%, which is 1.17% better than the best correct rate published by NIST; the correct rate on the Flickr 10K also reaches 79.27%.4. based on the improved algorithm of the depth confidence network tattoo image detection based on space Pyramid (SP-DBN). The algorithm uses the SPM model to solve the problem of the size and spatial distribution of the tattoo image. The experimental results show that the correct rate of the NIST tattoo data set is 97.23%, which is 0.93% better than the best correct rate published by NIST; the correct rate on the Flickr 10K has reached the improved algorithm of the convolution neural network image detection based on the three channel fusion of 80.46%.5.. CFT-CNN). In view of the feature extraction ability of all connected layers at different scales, a simple single channel T-CNN model is designed according to the detection of tattoo image. On the basis of single channel T-CNN model, a convolution neural network model (CFT-CNN) for three channel connection layer is designed, and the task is applied to the task of tattoo image detection. In the NIST data set, the correct rate of CFT-CNN is 97.87%, which is 1.57% higher than the best result published by NIST, and the correct rate on the Flickr 10K data set also reaches the 85.61%.6. based Faster R-CNN tattoo pattern detection improvement algorithm based on the three channel fusion (CFT Faster RCNN). In view of the size of the tattoo image and the large scale change, on the basis of Faster R-CNN, we fully consider the feature extraction ability of the full connection layer at different scales, and add a three channel full connection layer after the ROI pool, which solves the problem of the difficult position recognition of the tattoo image. The experimental results show that this calculation is based on the NIST data set. The method is compared with the Faster R-CNN algorithm on MAP, which is 3.35%, the 4.59%. is improved on IOU, and the improved algorithm based on DBN and CNN is compared. The results show that the improved algorithm improves the performance on NIST, and the improved CNN algorithm has better effect. In the improved DBN method, the word packet model is in SIFT. On the basis of features, the image of tattoo is expressed. On the small data set, the better.SP-DBN algorithm has a better effect on the larger data sets because of the consideration of spatial information. From the experimental results of two data sets of NIST and Flickr, the NIST dataset has some limitations, and the Flickr data set is closer to the real environment.

【学位授予单位】：南昌大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】