Research On Key Techniques Of Surface Defect Image Detection For Cables In Cable-stayed Bridges

With the rapid development of bridge construction, long-span and super long-spancable-stayed bridges and suspension bridges were widely used. Cables are the mainstressed part in these bridges and their reliability and durability are directly related tothe safety and service life. The cables are packaged by polyethylene (PE) or highdensity polyethylene (HDPE) protective layer outside. However, the cables suffer fromthe long-term alternating loads, and they are exposed to the natural environment. Thecables suffer from corrosions, affecting the service life span and resulting in bridgesafety accidents. Therefore, the cable surface defect detection is of great significance. Atpresent, methods for cable surface defect detection are mainly based on artificial visualdetection and laser scanning. These methods have many limitations such as lowefficiency, high cost, and lack of intelligence. According to the urgent need in this field,there has been in-depth study of related theory and key technology of machine visionsystem in this paper. A distributed machine vision system for the bridge cable surfacedefect detection and defect image recognition are proposed. In view of the collectedblur defect images, we put forward a blind image restoration method based onnon-negative domain support recursive inverse filter (NAS-RIF) and adaptive totalvariation (TV) regularization. In order to quickly and efficiently get the complete cablesurface defects, we proposed an improved scale invariant feature transform （SIFT）feature matching algorithm for automatic stitching. Finally, particle swarm optimization(PSO) algorithm is adopted to optimize the support vector machine（SVM）model. Thismodel is employed to classify the cable surface defects. The main research contents andcontributions are as follows.1）A distributed machine vision system for the bridge cable surface defect detectionand defect image recognition methods were proposed. Firstly, the climb robot loadingdistributed image sensors, a light source, an embedded DSP hardware platform, positionsensors, storage devices, four CCD image sensors distributed around the cable wereused to obtain the cable surface images. Then TI high-performance DSPTMS320DM642(DM642) was employed as the core processor to achieve real-timedefect image preprocessing, defect object segmentation, defect preliminaryidentification, and preliminary identification of defects in storage. Finally, imagedeblurring and image stitching were processed on peasonal computer (PC) to achieve the defect image recognition.2）The real-time realization methods of surface defect image preprocessing anddefect target extraction have been studied. By Analyzing the noise types and sources ofsurface defect images, an improved median filtering method based on the characteristicsof the embedded processor cache of DM642was proposed for cable surface imagefiltering processing. To realize rapid real-time preliminary identification of defect target,mathematical morphology (MM) and improved Sobel-edge-detection-algorithmcombined to a MM-Sobel image segmentation method was proposed and adopted toextract defect targets of defect images. Finally, the preliminary identification of thedefect was fulfilled. Moreover, defect images and their location information werestored.3）In view of blur surface defect images collected by the machine vision inspectionsystem, the blur models were analyzed. Thus, a blind image restoration algorithm whichcombined NAS-RIF and adaptive TV regularization was proposed. In this algorithm, wecombined with the mechanism of image degradation and image blind restorationmethods. In view of the original NAS-RIF algorithm being sensitive to noise problemsunder low signal-to-noise ratio, we added TV regularization constraint item to the costfunction of the original NAS-RIF algorithm. In order to effectively achieve the balancebetween image detail recovery and noise suppression, the total variation regularizationparameter was adjusted adaptively by maximum a posteriori probability. TheMajorization-minimization (MM) method and conjugate gradient iterative algorithmwere adopted to improve the rate of convergence of the algorithm. The experimentalresults indicated that the algorithm has good adaptability and effectiveness.4）The inspection system used4CCD cameras located around the cable to obtainthe cable surface images, a surface defect may be distributed in several images. In orderto identify the complete defect, it need automatic stitching on the corresponding defectimages. Image matching is a crucial step in image mosaic algorithm. According to thecharacteristics of the system for cable surface image, an improved SIFT featurematching algorithm was proposed. First, simple and effective Harris operator was usedto extract feature points. Then, according to the characteristics of the collected defectimages, the main direction assignment of the feature point and the rotation of thematching images were simplified in SIFT algorithm. Finally, the complete defect imagewas obtained by fusing the matching images. The experimental results show that thismethod greatly reduces the complexity of the SIFT algorithm and it can quickly and efficiently get the complete cable surface defects.5）Longitudinal cracks, transverse cracks, surface erosion, and scarring pit holesare the mainly categories scar defects in cable surface. Feature extraction and supportvector machine (SVM)are employed to classify these defects. In order to improve theSVM classification recognition rate, this paper uses the particle swarm optimization(PSO) algorithm to optimize the punish coefficient c and the kernel functionparameter g of the SVM model, namely the PSO-SVM algorithm. Through theexperiments for surface defect classification and recognition, classification recognitionrate reaches96.25%, the results show that PSO-SVM has high recognition rate and fastrecognition speed.The thesis focused on theoretical and experimental studies of image detection keytechniques for the bridge cable surface defect. The real-time performance of themachine vision inspection system was improved by using effective defect imagepreprocessing, defect object segmentation and defect preliminary discriminationalgorithms. To explore the applicable image deblurring algorithm, image stitchingmethod and classification recognition algorithm, the effect of the cable surface defectrecognition was also improved. In this thesis, the research on surface damage detectionand maintenance for the bridge cables based on machine vision detection system was ofgreat significance.