Research On Crack Detection Of Magnetic Ring Based On Machine Vision

This thesis presented a novel crack detection method based on crack mechanism analysis and image charateristics.In this study,an analysis of texture on the surface of magnetic ring and pseudo-crack feature is conducted before the study of filtering algorithm.The performance of edge-preserving filter towards the processing of the image of magnetic ring is studied.For that,several state-of-the-art edge-preserving filters,including bilateral filter,guide filter and weight least square filter,are discussed.Then a compartive experimental test computed by peak-signal-noise-ratio is executed to show the performance of different edge-preserving filters.An image sampling algorithm based on the polar coordinate transformation is proposed to reduce the raw data of crack detecting algorithm.Experiment show that it performs excellent in robustness.The mechanism of cracks is analyzed,according to which we establish the mathmatical model of cracks feature on the image.Combine the concept of gradient vector field with local descriptor,we propose a novel local descriptor based on image gradient vector(LDGV),which is proved effective in recognition of the pixel in crack region.According to the seed pixels detect by the LDGV,a front propegation based on local energy(FPLE)is proposed to search for the entire crack skeletion.With respect to the condition of the termination of searching,a stop criterion is proposed based on momentum method,by which we can aquire the whole crack skeleton with only a single seed pixel near the crack region.On the basis of the resulting skeleton,a dynamic thresholding method based on Otsu’ algorithm is proposed to precisely segment the crack region.Experiments prove it outstanding.A study of pattern regonition is conducted to distinguish between crack and interference.Specifically,for the feature extration of cracks,this study presents a new statistical feature called the standard deviation of residual(SDR)of the connected components,.What’s more,linear discriminant analysis is used to classify the crack regions and the non-crack regions.To reach the optimal classification result,we study the different combinations of the 13 features,then a training process and a testing are conducted.Experiment shows that the regonition rate reach to94.1%.