Research On Image Preprocessing And Classification Recognition Method Of Catenary Components In High-speed Railway

The working environments of high-speed railway catenary are complex.If it breaks down,it will affect the operating safety.Therefore,the working state of catenary components should be detected to ensure that these components operate correctly.With the rapid development of high-speed railway technology,the traditional detection methods of catenary components can not meet the requirements of accuracy,efficiency and stability.In this thesis,a non-contact detection strategy based on image processing technology is used to detect the working state of catenary components.As some catenary images acquired from WuhanGuangzhou high-speed railway have the characteristics such as noise and low contrast.Firstly,the system preprocesses these catenary images by denoising and enhancement.Then,it finishes the image segmentation and feature extraction.And finally,the catenary components are classified and recognized.So as to automatically detect the working state of catenary components,and provide a reliable guarantee for the safe operation of catenary.Firstly,in the aspect of catenary image denoising,aiming at the problem that some catenary images present noise affected by external natural environments,dust in tunnel and other factors,a lifting wavelet is proposed to replace the Laplacian transform in Contourlet transform for reducing the redundancy.Then it is combined with the cycle translation algorithm to denoise the added noisy images and the actual acquired noisy images.The experimental results show that,considering the visual effect of image denoising and the evaluation indexes of denoising effect,the proposed method has strong ability to suppress the noise of catenary images,and it has good denoising effect.Secondly,in the aspect of catenary image enhancement,some catenary images present the low contrast due to the influence of light intensity and external environments,and some catenary components are blurred or even invisible.These images will be enhanced by the combination of nonsubsampled Contourlet transform with the adaptive enhancement function.Specifically,aiming at the problem that different filter banks in nonsubsampled Contourlet transform have different enhancement results,the thesis proposes a method to determine an optimal filter banks for low contrast catenary image enhancement.The experimental results show that the proposed method not only makes the catenary components such as messenger wire holder,insulator,support,sleeve double ears,positioning support,return line and other related components clear and visible,but also protects the image details well,and then reduces the system missing detection of the catenary components.Thirdly,in the aspect of catenary image segmentation,in order to segment the catenary components accurately,aiming at the problem that too many parameters of traditional pulse coupled neural network are difficult to be determined,a simplified pulse coupled neural network is constructed by simplifying its structure and parameters,and the minimum cross entropy algorithm is used to determine its partial parameter adaptively.The experimental results show that the proposed method not only segments the catenary components effectively,but also presents few over-segmentation,under-segmentation and missingsegmentation phenomena.In addition,as there are many components in the catenary image,the sum of squared differences method and correlation method are used to match the catenary components,respectively.Under different conditions,the advantages and disadvantages of the two matching methods are concluded.Fourth,in the aspect of feature extraction and dimension-reduced of catenary components,too many features will affect the detection accuracy and efficiency.A total of18-dimensional feature vectors such as appearance ratio,entropy,contrast,rectangularity and Hu invariant moment are extracted for classifying and recognizing the catenary components.The dimension-reduced methods include the multiple dimensional scaling method,isometric mapping method,landmark isometric mapping method,and linear local tangent space alignment method.They are used to reduce the feature dimensions of catenary components,respectively.The experimental results prove that the optimal 9-dimensional features chosen by the linear local tangent space alignment method are suitable for the input of extreme learning machine.Then,in the aspect of classification and recognition of catenary components,as the result of extreme learning machine is easily affected by its parameters,the influences of hidden layer neuron number,activation function and other parameters on the extreme learning machine are discussed.Meanwhile,the corresponding relationships between the input variables and different models,the accuracy and efficiency of different models are discussed.Besides,a linear local tangent space alignment method combined with the kernel extreme learning machine is presented for classifying and identifying the catenary components.The results prove that,compared with the extreme learning machine model,the proposed method can improve the accuracy,efficiency and stability of catenary components classification and recognition.Finally,in order to reduce the impact of feature extraction on classification and recognition results,and further improve its accuracy and stability,the stacked auto encoder network,convolution neural network and deep belief network are used to analyze and process the catenary components.These components include messenger wire holder,positioning support,flat cantilever,insulator,42 type sleeve double ears,55 type sleeve double ears and rotating double ears.The experimental results show that the deep belief network can improve the accuracy and stability of catenary components classification and recognition.In summary,as some catenary images acquired from Wuhan-Guangzhou high-speed railway have the characteristics such as noise and extremely low contrast,they are preprocessed by image denoising and enhancement firstly.Then they are segmented,the features are extracted,the linear local tangent space arrangement-kernel extreme learning machine and deep belief network are used to classify and recognize these catenary components with high accuracy and stability.