图像识别算法在细胞筛查及火灾探测中的研究

发布时间：2018-09-03 16:41

【摘要】：图像识别主要包括图像预处理、特征提取、分类识别等内容。深入挖掘图像中的特征信息,提高特征提取的可靠度,是图像识别的关键环节。图像识别可分为静态目标识别与运动目标识别。静态目标识别可以通过单帧图像分析来完成,而动态目标识别通常需要依据视频或者图像序列的时间、空间连续变化规律来实现。无论是静态目标识别还是动态目标识别,都要结合图像自身的特点及识别目标的相关知识,才能设计出具有更高识别率及更好实时性的图像识别系统。宫颈细胞及火灾烟雾识别是当今学术界关注的两类热点问题。以静态的宫颈细胞及动态的火灾烟雾作为研究对象,在深入分析细胞及火灾烟雾图像特征的基础上,设计了具有良好性能的宫颈细胞筛查系统及火灾烟雾探测系统。主要工作如下:首先,对于较复杂的静态目标,图像分割耗时长、难度大。为了解决这一问题,对HE染色的宫颈细胞特性进行深入地分析,对比不同区块图像的特征,发现不同种类的块图像的纹理、颜色等特性呈现出较大差异,提出了用分块处理代替图像分割的思想,在此基础上,设计了一种基于块图像特性分析的宫颈细胞筛查系统。利用背景区块与非背景区块图像的特征差异,去除背景区块,减少后续处理量;随后,将非背景块分为正常块、异常块,对两种块图像的特征进行深入分析,并通过特征选择,提取出11个特征,通过T检验(p0.001)验证了所选特征具有较大区分度;根据所选特征设计支持向量机结构及参数,实现正常、异常块图像的分类识别。通过对比实验证明,本文所设计的基于块图像特征的宫颈细胞识别系统,相较于基于分割的识别系统,在保证准确率的基础上,提高了系统的实时性,为临床应用提供了更多可能性。其次,常规的图像识别系统,通常需要人工选择识别对象的特征,使得系统具有较强的局限性,一般只对某种特定环境下的特定对象有很好效果。就宫颈细胞而言,已有的识别系统,通常只对某一种染色下的宫颈细胞效果较好,而对于不同染色下的宫颈细胞,需要设计不同的识别系统。针对这一问题,利用卷积神经网络的特征学习能力及支持向量机的分类优势,对卷积神经网络进行改进,用支持向量机代替传统卷积神经网络的输出层,构建了一个基于支持向量机的卷积神经网络。针对不同染色的宫颈细胞这一研究对象,设计网络的相关参数,实现正常、异常细胞的分类识别。通过对比实验证明,本文提出的识别系统比常规的卷积神经网络及支持向量机准确率更高。最后,仅采用静态特征模型不能描述目标的动态特性,且很难解决动态目标场景复杂、快速移动等问题。因此,对于运动目标的识别,需要结合静态、动态特性,综合分析。以火灾烟雾为研究对象,为了实现极早期火灾探测,利用烟雾特殊的湍流特性,对可疑运动区域轮廓的光流矢量特性进行提取,代替传统的对于整个可疑区域光流矢量的计算。在此基础上,提出了一套“动-静-动”的识别方案:先利用改进的帧间差分法检测出运动目标,排除静态物体的干扰;随后,通过检测运动区域的颜色模型及纹理特征,排除部分运动干扰;采用基于高斯金字塔的Lucas-Kanade光流法,分析运动区域轮廓的光流矢量,排除非烟干扰,确定烟雾区域。通过对比实验证明,本文提出的烟雾识别方案不仅保证了准确率,也提高了实时性。
[Abstract]:Image recognition mainly includes image preprocessing, feature extraction, classification and recognition. Deep mining of image feature information to improve the reliability of feature extraction is the key link of image recognition. Image recognition can be divided into static object recognition and moving object recognition. In order to design an image recognition system with higher recognition rate and better real-time performance, both static and dynamic object recognition should be combined with the characteristics of the image itself and the relevant knowledge of the target recognition. Cell and fire smoke identification are two hot topics in academia nowadays. Taking static cervical cells and dynamic fire smoke as research objects, a cervical cell screening system with good performance and a fire smoke detection system are designed on the basis of in-depth analysis of cell and fire smoke image characteristics. In order to solve this problem, the characteristics of cervical cells stained with HE are analyzed in depth, and the features of different blocks are compared. It is found that the textures and colors of different kinds of blocks are quite different, and a block processing is proposed to replace the image. Based on the idea of image segmentation, a cervical cell screening system based on the analysis of the characteristics of the block image is designed. Eleven features are extracted by feature selection, and the T-test (p0.001) verifies that the selected features have a large degree of discrimination; support vector machine structure and parameters are designed according to the selected features to achieve normal and abnormal block image classification and recognition. Compared with segmentation-based recognition system, it improves the real-time performance of the system and provides more possibilities for clinical application on the basis of ensuring the accuracy. Secondly, the conventional image recognition system usually needs to select the features of the recognition objects manually, which makes the system have strong limitations, generally only for specific pairs in a specific environment. As far as cervical cells are concerned, the existing recognition systems are usually better for one kind of stained cervical cells, but for different stained cervical cells, different recognition systems need to be designed. A convolution neural network based on support vector machine is constructed by improving the product neural network and replacing the output layer of the traditional convolution neural network with support vector machine.The relevant parameters of the network are designed for different stained cervical cells to realize the classification and recognition of normal and abnormal cells. The proposed recognition system is more accurate than conventional convolution neural network and support vector machine. Finally, the static feature model alone can not describe the dynamic characteristics of the target, and it is difficult to solve the dynamic target scene complex, fast moving and other problems. Therefore, for moving target recognition, it is necessary to combine static and dynamic characteristics, comprehensive analysis. In order to detect fire haze in the very early stage, the optical flow vector characteristics of suspicious moving region contour are extracted by using the special turbulence characteristics of smoke, instead of the traditional calculation of optical flow vector for the whole suspicious region. The frame difference method based on Gauss pyramid is used to detect moving objects and eliminate the interference of static objects. Then, some motion disturbances are eliminated by detecting the color model and texture features of moving regions. The Lucas-Kanade optical flow method based on Gauss pyramid is used to analyze the optical flow vector of moving region contour, eliminate non-smoke disturbance and determine the smoke region. Experiments show that the smoke recognition scheme proposed in this paper not only guarantees accuracy, but also improves real-time performance.
【学位授予单位】：天津大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】