| Super smooth high gloss reflective surface products are widely used in engineering,for example,automotive paint film,bearings,solar silicon panels,etc.In the production process,due to the influence of material inclusions,dust and other factors in the production environment,there will be slag spots,hairlines,scratches,cracks,unevenness and other small defects on the surface of the product,such as dots,lines and surfaces,which affect the quality of the product.In the use of machine vision inspection,due to the reflection of high light on the surface,ambient light interference,resulting in defects missed and misjudgment.The detection of defects is a very important part of ensuring product quality in automated production,and the machine vision detection of tiny defects on ultra-smooth high-gloss reflective surfaces is a hot spot and a difficult area of research now.In this paper,we study the design of the vision inspection system,mainly including defect image acquisition,image pre-processing,defect detection and identification,with the application background of automatic inspection of automotive paint film defects.Firstly,the design of the lighting source is studied for the problem that the surface smoothness of the object is high,the light reflection is high,and the ambient light will affect the imaging quality of the defects,and the design of the lighting source is studied.To address the problem that microscopic defects are seriously disturbed by noise and background,a method is proposed to enhance the defect information by combining homomorphic filtering and wavelet transform,using superposition fusion noise reduction and homomorphic filtering to improve the contrast between microscopic defects and background,and then using two-dimensional wavelet transform to remove the background trend term to obtain more stable defect information;to address the problem of detecting microscopic defects,a method is proposed to detect and identify defects.Finally,an optimized AlexNet neural network model is proposed to accurately classify and identify the defective data set based on the segmentation results.The experimental results show that the acquisition system can obtain high-resolution defect images,and the processing method can achieve effective detection of small defects(0.1mm)on large-area ultra-smooth high-light reflective surfaces,and the detection rate of small defects reaches 97.6%,the false detection rate is 1.3%,and the accuracy of defect identification can reach 90.56%,which achieves the expected detection results. |
