复杂背景下牛体检测的研究与实现

发布时间：2018-03-02 11:10

本文关键词： 图像处理同态滤波图像分割牛体检测混淆矩阵　出处：《西北农林科技大学》2015年硕士论文　论文类型：学位论文

【摘要】：肉牛养殖业在我国的国民经济建设中占有非常重要的地位,是高效、节粮、与民生息息相关的产业,同时也是我国调整农业结构的战略方向。产肉量作为评定肉牛质量的重要指标,更是迫切需要实现自动化测定。牛体检测作为其一项前期准备工作,有着十分重要的意义,本文针对这一问题,对牛体图像进行分析处理,主要通过滤波处理、牛体检测、形态学运算以及评价对比等方法将牛体与复杂的自然背景分离,并获得了较好的分离效果。本文的主要研究内容有以下几个方面:(1)图像预处理:由于牛体图像在获取的过程中会受到光照影响,所以要对图像进行对比度增强和光照去除处理。本文首先在YCbCr空间上对牛体图像进行分解,得到其Y通道图像,然后用同态滤波对其进行滤波处理以增强图像对比度。但处理结果对环境光照影响去除效果不明显,因此对上述Y通道图像进行二级小波分解及同态滤波处理,使图像基本信息在最低分辨率层得到体现,且图像峰值信噪比达到了24.35。因此,在YCbCr空间Y通道下采用小波变换及同态滤波方法,可以降低牛体检测过程中强自然光照的影响,且增强了图像的对比度,为下一步牛体检测做好了准备工作。(2)牛体检测:本文通过两种方法来进行牛体检测,分别是贝叶斯分类器和改进大津算法,即分别利用皮肤检测和图像分割的方式将牛体从复杂的背景中分离开。贝叶斯分类器采用了RGB和HSV两个颜色空间实现牛体的检测,通过统计在不同颜色通道上的检测效果建立了相应的训练集和测试集。(3)图像优化及评价:牛体所处环境和其自身毛色均较为复杂导致牛体检测结果存在背景噪声,因此需对图像进行优化处理。本文运用形态学运算对牛体检测图像进行修复和优化,得到了较为完整的牛体信息。然后,采用混淆矩阵对优化后的图像进行评价和分析。通过对最终处理得到的20幅牛体图像进行分析和计算,结果表明,贝叶斯分类器和改进大津算法的牛体提取平均准确率分别为86.17%和80.07%。本实验较好的解决了牛体与复杂背景分离的问题,基本实现了牛体的完整提取,为后续牛体自动化测量提供前期准备工作。
[Abstract]:Beef cattle farming occupies a very important position in the national economic construction of our country. It is an industry that is efficient, grain saving and closely related to people's livelihood. At the same time, it is also the strategic direction of adjusting the agricultural structure of our country. As an important index to evaluate the quality of beef cattle, it is urgent to realize automatic determination. In order to solve this problem, this paper analyzes and processes the bovine body image, mainly separates the bovine body from the complex natural background by filtering, detecting, morphological operation and evaluation and comparison, etc. The main contents of this paper are as follows: image preprocessing: the bovine body image will be affected by light during the process of acquisition. Therefore, contrast enhancement and illumination removal should be carried out on the image. Firstly, the bovine body image is decomposed in YCbCr space, and the Y channel image is obtained. Then the homomorphic filter is used to filter the image to enhance the contrast of the image. However, the effect of the processing result on the environmental illumination is not obvious, so the Y-channel image is processed by two-level wavelet decomposition and homomorphic filtering. The basic information of the image is reflected in the lowest resolution layer, and the peak signal-to-noise ratio (PSNR) of the image reaches 24.35.Therefore, wavelet transform and homomorphic filtering in YCbCr space Y channel can reduce the influence of strong natural illumination in the process of bovine body detection. And enhanced the contrast of the image, prepared for the next step of bovine body detection. 2) Bovine body detection: this paper through two methods to carry out cattle body detection, respectively, Bayesian classifier and improved Otsu algorithm, That is to say, cattle body is separated from complex background by skin detection and image segmentation. Bayesian classifier uses two color spaces, RGB and HSV, to detect bovine body. The image optimization and evaluation of training set and test set on different color channels were established by statistical analysis. The results showed that the environment of cattle body and its coat color were more complex, which resulted in the background noise in the result of bovine body detection. Therefore, it is necessary to optimize the image processing. In this paper, the image of bovine body detection is repaired and optimized by morphological operation, and the complete information of bovine body is obtained. The confusion matrix is used to evaluate and analyze the optimized images. The analysis and calculation of 20 bovine body images obtained from the final processing show that, The average accuracy of bovine body extraction by Bayesian classifier and improved Otsu algorithm is 86.17% and 80.07 respectively. This experiment has solved the problem of separating cattle body from complex background and basically realized the complete extraction of bovine body. To provide early preparation for follow-up automatic measurement of cattle body.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：S823;TP391.41

【参考文献】