| As an indispensable part of the exterior surface of the body,the paint film protects and decorates the body.Due to the influence of factors such as environment,coating process and coating quality of automobile coating workshop,it is inevitable that various kinds of defects will be formed on the surface of paint film.Aiming at the problems of low efficiency and unstable detection results of traditional manual visual inspection,this thesis combined the advantages of traditional machine learning defect detection algorithm on image bottom texture feature extraction with the advantages of deep learning defect detection algorithm on semantic feature extraction on the basis of full investigation of domestic and foreign paint film defect detection methods.A new paint film defect detection algorithm based on machine vision was proposed,which realized two stages of defect area rough detection and defect detection and recognition from rough to fine.The main contents are as follows:(1)In order to solve the problem of single background and small defect area of paint film sample,the sample image was segmented and the defect area was roughly detected by classifying the segmented molecular image.Firstly,the statistical features of gray histogram and gray co-occurrence matrix of the sample sub-image were extracted.Secondly,the fusion classification model of gradient lifting tree and logistic regression was built,and the key parameters were optimized by random search strategy.(2)Based on the model detection accuracy and detection speed,a paint film defect detection and recognition algorithm based on YOLOX-tiny was proposed,and the network structure of YOLOX-tiny was improved by integrating the convolutional block attention mechanism module for the large proportion of paint film defects and the extreme aspect ratio.(3)According to the surface characteristics of paint film sample provided by a company,an experimental platform for image acquisition of paint film sample was built.A sample set containing normal sub-images and defect sub-images was obtained by segmentation of sample images,and the number of defect sub-images was amplified by off-line data enhancement.At the same time,the defect target was marked in the defect sub-image,and the paint film defect data set was established to make data preparation for the rough detection of defect areas and the detection and recognition of defects.Finally,through the experimental comparative analysis of different classification models on the feature data extracted in this thesis,the classification accuracy of the classification model adopted in this paper reaches 94.24% on the test set,which is 1.43% and 2.87%higher than the GBDT and LR classification models alone.At the same time,for paint film defect detection and recognition,the YOLOX-tiny model based on CBAM attention mechanism proposed in this thesis has an average detection accuracy of 93.31%,which is2.09% higher than the model without attention mechanism,and the detection accuracy of tiny defects and extreme aspect ratio defects has also been greatly improved. |
