| Stress concentration and alternating load are inevitable accompanied with the manufacturing,transportation and application process,which may lead to harmful fatigue fracture during the long-term application.Due to different materials feature and complex object structure,various forms of external force,the larger dynamic range of the scene,hence the surface testing of the structure is a high difficulty work,which needs practical theory and has important social significance and economic value.Although the traditional crack testing technology involves in different fields of social production and has unique advantage in the various object of detection,most methods require operation in accordance with the specific testing and installation conditions.With the development of social economy and automation technology,computer vision and image processing algorithm have been gradually used in medical,criminal investigation,remote sensing,aerospace and other research fields.Since the 1960s,crack detection techniques based on image segmentation have been widely used in many industrial sectors including automotive manufacturing.Meanwhile,varieties of segmentation algorithms for different cracked research objects have been paid great attention in academics and a few of approaches have been proposed.This fact indeed contributes to promoting the practical application of image crack detection technology.However,since the complex working condition and diversity of crack are frequently encountered in the image crack detection research filed,which leads to a crucial yet non-tribal issue that should not be over-estimated.To be more specific,the problem can be divided into four types,i.e.,how to precisely extract the attribute characteristics of the crack;how to significantly improve the accuracy of the feature clustering;how to feasibly give the target model or classifier a strong optimization classification performance and how to properly make a trade-off between the accuracy and speed of the algorithm in the presence of complex and similar background.In this respect,how to propose a proper algorithm for crack image segmentation is still an open problem and deserves further investigating.From the comprehensive perspective of improving the robustness and reducing the error classification risk of crack image segmentation,in this paper,crack image segmentation under unconstrained scene is studied on the cracks existing in the road,wall and steel beam,we introduce the wavelet transform,multi-scale normalized cut,multi-scale structured forest and deep learning that based full convolutional network into the field of crack image segmentation.We further investigate several novel algorithms of crack image detection in the bases of above four methods.To be more specific,these methods conclude the following techniques:multi-scale normalized cut based on anti-symmetrical biorthogonal wavelet transform,the multi-scale down-sampled normalized cut based on anti-symmetrical biorthogonal wavelet transform,the fast crack edge detection using structured forests and wavelet transform,and the improved fully convolutional network method.The main contributions of this paper can be summarized as follows:The main research contents of this paper can be summarized as follows:1.A total of 1576 images including road,wall and steel crack images were collected in natural scenes,standard image data set of crack image in unconstrained scenes were manual annotated.We use this data set and the corresponding annotation results to validate the effectiveness of crack image segmentation method that proposed in this paper;2.From the perspective of effective extract properties of the crack edge,modulus maximum edge detection of the multi-scale wavelet is introduced into the image crack segmentation.The respectively sequence and two scales of modulus maximum in 3 kinds of wavelet bases and 5 kinds of wavelet series are carried out in the experiment of image segmentation with crack image.6 kinds of wavelet methods that possess better perform than that of 6 kinds of traditional methods are both used to edge detection.The over-all techniques are carried out to validate the preforms in terms of qualitative and quantitative evaluation approach.3.It must be noted that the obliterated edge information and complexity of construct process of similarity matrix are generally inevitable in the image processing by using the multi-scale normalized cut.By employing the semi-reconstructed properties of the anti-symmetric biorthogonal wavelet to find the local modulus maximum,we can effectively extract the edge feature of multiple scales in multi-scale normalized cut.Through reducing construction process of similarity matrix,constructing down-sampling similarity matrix and multi-scale normalized similarity matrix,we can improve similar clustering of crack feature on the premise of reducing time-consuming.In this paper,the crack images are selected as the target,various wavelet methods under the improved method,multi-scale normalized cut,and other methods based on multi scale normalized cut method carried out to validate the preforms in terms of qualitative and quantitative evaluation approach.4.In order to address the problems in which lower precision of crack image segmentation are encountered in the introduced structured forest,the semi-reconstructed modulus maximum edge detection of the anti-symmetric biorthogonal wavelet is developed in feature extraction of structured forest.Multiple crack images are applied as the training and validation object,which aims at constructing classifier of multi-scale structured forest.At the same time,the crack images,the single crack images and stitched crack images of the steel beam are selected as the target,various wavelet methods under the improved method,multi-scale structured forest,and other methods based on multi scale normalized cut method in the original scale and multi-scale carried out to validate the preforms in terms of qualitative and quantitative evaluation approach.5.In general,regarding the fully convolutional network,where if object is smaller than the receptive field that fragment and mislabeling might be encountered.Furthermore,small objects are frequently neglected and classified as background and fully convolutional network is merely good at extracting the overall shape of an object.Given that the mentioned discussion above,we will propose the changed network structure of the fully convolutional network by means of increasing the convolution layer,Relu technology,deconvolution layer and remove the Dropout layer.Multiple crack images are applied as the training and validation object,which aims at comparing the model of fully convolutional network and the improved network models with the qualitative and quantitative evaluation.At the same time,the evaluation of the 4 segmentation methods proposed in this paper is presented by using the crack image data set in unconstrained scene.The results show that the proposed Crack FCN method can significantly improve the robustness and reduce the misclassification risk of the crack image segmentation in the unconstrained scene. |
PDF Full Text Request