Research On Automatic Detection Methods Of Rail Surface Defects Based On Image Processing

Automatic detection of rail surface defects is an important way to ensure security of railway operation. Because of its accuracy, high efficiency and high degree of automation, the method using image processing to detect rail surface defects has become an important research topic. On the basis of previous research results, this paper deeply researches on several key technologies involved in automatic detection of rail surface defects, such as image denoising, defects edge detection, defects segmentation, defects classification. Main work is as follows:Firstly, a denoising method for rail surface defects image based on anisotropic diffusion in shearlet domain is proposed. The noisy image is decomposed by NSST. Then, K-SVD algorithm is used to remove noise in low frequency component, and KAD algorithm is used for noise reduction of each high frequency component. Finally, the reconstruction image is obtained by INSST for the processed low frequency component and high frequency components. Experimental results show that, compared with denoising methods such as wavelet combining with nonlinear diffusion method, shearlet hard threshold method, K-SVD sparse denoising method and sparse redundant denoising method in wavelet domain, the proposed method has a better performance in noise reduction and can preserve the detail and textural features of rail surface defects more efficiently.Then, a denoising method for rail surface defects image using kernel fuzzy clustering and regularization is studied. Firstly, KFCM clustering algorithm is used for clustering the similar image pieces. Then, a norm regularization constraint condition is imposed and sparse decomposition of image pieces in the same class under the dictionary is achieved. Finally, the update of dictionary is completed by improved K-SVD algorithm, thus, noise in the image is suppressed effectively. Experimental results show that, compared with denoising method based on wavelet combining with nonlinear diffusion, denoising method based on constant dictionary, denoising method of optimal directions and K-SVD denoising method, the proposed method can remove noise of the image more effectively and preserve the details of rail surface defects and improve the visual effect better.Next, a method of edge detection based on improved bee colony in shearlet domain is given. Rail surface defects image is decomposed by NSST. Then, the improved bee colony algorithm is used to detect the basic contour line of image edge of low frequency component accurately, while the direction modulus maxima algorithm is applied to high frequency components, thus the abundant details of image edge can be detected. Finally, the detection results of low frequency component and high frequency components are fused. Experimental results show that, compared with the existing methods of edge detection such as Canny method, improved ant colony method, improved bee colony method, improved NSCT modulus maxima method, the image edges detected by the discussed method are located accurately and can be complete and clear, with abundant details. The method can more effectively improve the performance of edge detection and requires less running time.And then, a method of dual-threshold segmentation for rail surface defect image using Arimoto entropy based on chaotic bee colony optimization is given. The method of single-threshold selection based on Arimoto entropy is extended to dual-threshold selection. Then intermediate variables in formulae of Arimoto entropy dual-threshold selection is calculated by recursion to reduce the amount of calculation. Finally, bee colony algorithm is improved by chaotic sequence based on tent mapping to achieve the fast search for two optimal thresholds. A large number of experimental results show that, compared with the existing segmentation methods such as multi-threshold segmentation method using maximum Shannon entropy, two-dimensional Shannon entropy segmentation method, two-dimensional Tsallis gray entropy segmentation method and multi-threshold segmentation method using reciprocal gray entropy, the given method can segment rail surface defects more quickly and accurately with superior segmentation effect. It proves to be a real-time and effective detection method for rail surface defects.Finally, a classification method of rail surface defects based on sparse representation and random forest is proposed. Firstly, the sparse dictionary that is obtained by using K-singular value decomposition algorithm to train the sample images is used to decompose rail surface defect images. Then the feature vectors are constructed by decomposition coefficients and the dimension of feature vectors is reduced by principal component analysis algorithm to ensure the efficiency. Finally, feature vectors are trained and tested by random forest classifier. A large number of experimental results show that, compared with the method based on shape and gray characteristics and BP neural network, the method based on Gabor and SVM and the method based on sparse representation and SVM, the proposed method can express the image characteristics more effectively than other methods. It attains the highest classification accuracy and the fastest classification speed.