| Al-alloy material is often used as the material of choice for notebook computer panels,but the surface may be damaged due to external factors during the processing of the panel.Therefore,surface quality inspection is an indispensable and important part of the entire production process.At present,the detection of surface defects of notebook Al-alloy panels in the industry mainly relies on human eye detection,and the size of defects that can be detected by human eyes is generally large.For small defects,that is,the size is 0.05mm^2 and below,the depth is shallow or the shape is thin and long,the human eye detection is difficult to meet the requirements.The surface defect detection method based on machine vision often has the problem of insignificant imaging of small defects in the industry.That is,the collected defect images often have problems such as low contrast between the defect target and the background,and the number of pixels represented by the defect,which makes it difficult to obtain features.Highlighted high-quality defect images.At the same time,in the subsequent defect detection process,the problem of missing defect detection and false detection will occur,that is,the detection method will miss the true defect,and the non-defect or false defect will be judged as true.Aiming at the problem of insignificant imaging of small defects,this paper studies the photometric three-dimensional surface defect measurement method combining symmetrical light source and spectroscopic correction.It is suitable for the imaging of flat animals at a uniform speed and can quickly and stably measure the gradient distribution of the defect surface.At the same time,this paper designs a high-quality defect image generation method based on multi-channel Gaussian fusion.This method combines the curl,divergence,Gaussian curvature and average curvature of the gradient distribution,so that a single image has more feature information and highlights the defects.It also uses FFT-based Gaussian filtering method.Compared with ordinary Gaussian filtering method,the fusion efficiency is increased by 5.15 times,so that the imaging method in this paper also has a higher calculation speed at high resolution.Aiming at the problem of missing defect detection and false detection,this paper designs a defect detection method combining multi-scale feature fusion and cascaded residual network.This method is based on a one-stage target detection framework,combined with a cosine annealing and hot restart random small batch gradient descent method,So that the multi-task loss function can effectively converge in the iterative process.Backbone in this method uses Cascade-DResnet,a cascaded residual network that combines random subspace RSM and depth separable convolution,to verify that the classification accuracy is 97.66% in the test set,and quantize the original float32 model through the model quantification method For the int8 model,the model compression rate is 46.9%.With almost lossless accuracy,the FPS reaches 78,and 1.77 panels can be identified per second.Finally,an experimental platform was built according to the design scheme of the surface defect optical measurement system,the measurement method and defect imaging method designed in this paper were integrated into the imaging module,the defect detection method was integrated into the target detection module,and the software application system was designed and completed.Used for interactive operation,production testing and network model training.5572 images were collected through the experimental platform.The experimental results show that m AP exceeds90%,and the detection speed reaches 69 frames per second,which basically meets the production requirements of real-time online detection. |
PDF Full Text Request