Study On Fast Detection Method Of Defects On Automobile Painting Surface

With the rapid development of our country, family car ownership inChina is gradually increasing. Automobile painting is a main reflection ofits appearance, and a clean, tidy and flawless car surface will surely leavea great impression to a car buyer. Nowadays, customers demand forsafety and comfort as usual, meanwhile appearance of a car is playing amore and more important role in people’s eyes. Cars will be checkedthoroughly before they go to the market in both domestic and overseasautomobile plants.However, automatic detection of car painting defects is not widelyused so far, for the reason that the characteristics of defects are not soclear and the surface structure of a car is complicated, most of theautomobile factories choose to finish this step by manual work. Theefficiency of manpower is not so acceptable, and the labor cost is too high. Besides, the detect results might be affected by workers’ experience,physical and mental condition or fatigue degree. In recent years,computer technology, digital image processing technology, and automaticcontrol technology have achieved rapid progress. All of these have laid agood technical foundation for automobile painting defect detection.With the support of SVW (Shanghai Volkswagen Automotive Co.,Ltd), this article analyzed the characteristics of automobile paintingdefects, and proposed a low-angle lighting solution from four directionsby combining the characteristics of defects and the practical applicationbackground, by this means, clear images of painting defects can be got.Using FFT (fast Fourier transform), DoG (Difference of Gaussians) filterand gray level transformation to process images so that the defect areascan be more obvious. At last, a modified contour extraction algorithm isused to obtain the defect area contour, experiment results showed that thecorrect rate of this algorithm is beyond80%. In addition, to realizedefects classification, this paper came up with a classification algorithmbased on modified Local Binary Pattern (LBP), training a large number ofsamples’ LBP feature data to get some specific classifiers. Experimentswere conducted specific to five certain kinds of defect, namely particle,fisheye, scratch, popping, exotic matter. SVW has provided us64painting samples with16colors and1roof of a real car. The resultsshowed that the effective rate of this classification method on these typical defects can be above91%and the correct rate is more than82%,processing time of one single image is less than60ms.