Part Segmentation Marking And Defect Detection Based On Computer Vision

Auto parts are an important cornerstone for the healthy and sustainable development of the automotive industry.If a car with defective parts is flowed into the market,it may cause serious traffic accidents and even harm the safety of people’s lives and property.Parts are rigorously screened during the production phase of the automotive industry and defective parts are rejected.However,the original manual screening is not only inefficient and inaccurate,but also consumes a lot of manpower and material resources.Part defect detection based on computer vision came into being,aiming to achieve an efficient leap from "human inspection" to "machine inspection".At present,there are still many problems in the field of computer vision in the field of part defect detection.The first is that the number of part defect samples is small,and the use of complex neural networks to train on small data sets is prone to overfitting,and the poor detection.In addition,the number of parts is huge,but the defects of parts are very small,which has certain requirements for the details and detection speed of the network.There are also some models that only have good effect on some individual data sets,poor adaptability,and the object detection algorithm is mostly artificial box labeling,which cannot be specific to the specific pixels of the target.In view of the above problems and research background,this paper carries out the following work:(1)Based on the construction of part defect data set,a kind of image segmentation annotation to improve U-Net is proposed algorithm.Based on the infrastructure of the U-Net model encoder and decoder,the algorithm makes a series of related improvements to the characteristics of automotive part defects to improve the segmentation accuracy of the network.The main improvement points are as follows: 1)In the encoder part of U-Net,Res2 Net is used instead of the original convolutional structure,and the multi-scale features are represented in a more fine-grained manner by increasing the residual block,thereby increasing the size of each layer of the sensing field,reducing the loss of image information,and further improving the accuracy of the image segmentation task.2)Increase the convolution of holes between the encoder and the decoder,expand the model sensing field without changing the feature map size,and reduce the loss of image detail during the down sampling process.3)Connect a Mini U-Net to the output part of the decoder to re-patch the model to solve the blurring and loss of detail caused by mostly minor defects on the part defects.(2)According to the characteristics of automotive part defects,a part defect inspection based on the ECA-SSD model is proposed algorithms.The model is based on the basic architecture of the SSD model and makes the following improvements: 1)Replace the traditional convolutional structure of the VGG-16 part of the SSD model with deep separable convolution,which greatly reduces the model parameters and improves the detection speed.2)Use the linear bottleneck inverted residual structure,and appropriately increase the network structure without increasing the amount of computation to improve the accuracy of inspection.3)Replace the Re LU activation function with the Re LU6 activation function to avoid the loss of information caused by setting unlimited.4)Increase the ECA-Net effective channel attention mechanism,while avoiding dimensionality reduction,the model can focus on the target of the image,while ignoring the impact and interference caused by the background.And the feature layer is weighted and activated,and different attention levels are set for different feature layers.So using this method to improve the detection accuracy of the model.(3)Comprehensive use of improved U-Net split annotation + ECA-SSD model for part defect detection.The general box annotation of defect targets has been improved to pixel-specific segmented annotations,making the results of defect detection clearer and more specific,more suitable for multi-target detection scenarios.Experimental results show that in the segmentation annotation stage,the size of the segmentation model in this paper is about 13.9M,the accuracy rate reaches 84.61%,and the time is 0.017 s.The task of segmenting part defects can be better completed.In the target defect detection stage,the model size is less than 15 M,the accuracy rate reaches more than 93%,the time is 0.034 s,and the detection effect is good.Meantime,the model has the characteristics of lightweight and pixelated,which meets the requirements of industrial production for the speed and accuracy of part detection.