【作者】 高强；

【导师】 鲁建华；

【作者基本信息】 大连理工大学 ， 船舶与海洋工程， 2020， 硕士

【摘要】 依赖声呐设备的水下目标分类识别技术,在海洋资源勘探、水下鱼类识别、水下打捞等环境中扮演着越来越重要的角色。水下目标识别的一般步骤为原始声呐图像获取,图像预处理,图像特征提取,目标分类识别四个步骤。在声呐图像目标识别的整个过程中,每一个环节都有着决定性作用。其中,特征提取环节尤为重要,特征的不变性,即位移、尺度和旋转不变性越好,以及抗噪性能越强,其分类识别能力也越高。基于小波不变矩特征提取算法的工作已经有了很多,但是前人基本上都是将其应用在光学图像。在声学图像层面,尤其是水下目标识别这一领域研究尚少。针对本试验特定的扇贝和海星目标,考虑到其声呐图像轮廓模糊、噪点多的缺陷,将Hu矩、Zernike矩和小波不变矩应用在该类图像上,主要内容和成果包括:(1)详细介绍了Hu矩、Zernike矩和小波不变矩特征的基本概念和原理,对算法的位移、尺度和旋转不变性进行了推导与证明。(2)声呐图像矩特征提取试验。以扇贝为研究对象,对原始声呐图像分别作位移、尺度和旋转变换,并加入不同程度的高斯噪声。使用MATLAB进行编程提取各种矩算法下的特征值,分析对比不同算法下特征的不变性及其抗噪性能。(3)基于BP神经网络的海星和扇贝分类识别试验。采集3000幅原始声呐图像,对其做预处理和特征提取,分别将三种不变矩算法提取的特征值进行特征选择以后,输入神经网络分类器,通过识别效果来评判三种不变矩算法的性能。本文的创新点总结如下:(1)对于本文的扇贝和海星声呐图像,在传统的全局矩Hu矩和Zernike矩基础上,提出了兼具局部特征和抗噪性能的小波不变矩,并通过试验数值验证了小波不变矩强大的适应性能。(2)为了取得更好的识别结果,提出了一种基于特征融合和特征降维的方法,试验结果表明,该方法提高了整体的识别率,同时降低了计算时间。更多还原

【Abstract】 The underwater target classification and recognition technology relying on sonar equipment plays an increasingly important role in marine resource exploration,underwater fish identification,underwater salvage and other environments.The general steps of underwater target recognition are original sonar image acquisition,image preprocessing,image feature extraction,and target classification recognition.In the entire process of sonar image target recognition,each link has a decisive role.Among them,the feature extraction link is particularly important.The invariance of features,that is,the better the invariance of displacement,scale and rotation,and the stronger the anti-noise performance,the higher the classification and recognition ability.There has been a lot of work on the feature extraction algorithm based on wavelet invariant moments,but the predecessors basically applied it to optical images,acoustic images,especially underwater target recognition.For the specific scallop and starfish targets in this experiment,considering the defects of the sonar image with blurred contours and many noises,the Hu moment,Zernike moment and wavelet invariant moment are applied to this type of image.The main content and results include:(1)The basic concepts and principles of Hu moment,Zernike moment and wavelet invariant moment characteristics are introduced in detail,and the displacement,scale and rotation invariance of the algorithm are derived and proved.(2)Sonar image moment feature extraction experiment.Taking scallops as the research object,the original sonar image was transformed by displacement,scale and rotation,and Gaussian noise of different degrees was added.Use MATLAB to program and extract the eigenvalues under various moment algorithms,and analyze and compare the invariance and anti-noise performance of the features under different algorithms.(3)Starfish and scallop classification and recognition test based on BP neural network.Collect 3000 original sonar images,perform preprocessing and feature extraction on them,and select the feature values extracted by the three invariant moment algorithms respectively,and then input them into the neural network classifier to judge the three invariant moment algorithms through the recognition effect Performance.The innovations of this article are summarized as follows:(1)For the scallop and starfish sonar images in this paper,on the basis of the traditional global moment Hu moment and Zernike moment,a wavelet invariant moment with local characteristics and good noise resistance performance is proposed,and the wavelet is not verified by experimental values.Variable torque with strong adaptability.(2)In order to obtain better recognition results,a method based on feature fusion and feature dimensionality reduction is proposed.Experiments show that this method improves the overall recognition rate and reduces the calculation time.更多还原