| In the production process of mechanical products,due to factors such as technology and processing equipment,the surface of the product will inevitably produce defects.In order to ensure the quality and consistency of the product,surface defect detection is indispensable in the quality inspection process of the product of.Most of the surface defect detection of existing industrial panels still adopts manual mode,which is low in efficiency and high in cost.With the development of image processing technology,defect detection algorithms based on machine vision have been favored by researchers at home and abroad.Traditional defect detection algorithms rely on artificial design and expression of defect characteristics,and the established defect detection algorithm model has low robustness and weak generalization,and its detection performance is not good when the background features are similar and the size of the defect features is too different.In recent years,convolutional neural networks in machine vision algorithms have developed in-depth in target detection and other fields,gradually breaking the bottleneck of traditional defect detection algorithms,making up for the deficiencies of manual quality inspection,and achieving good results.On this basis,this paper proposes an industrial panel defect detection algorithm based on machine vision.The research contents are as follows:(1)This paper analyzes the performance of different target detection algorithms and the morphological characteristics of defect samples,and establishes an optimization route based on the Faster R-CNN algorithm.Aiming at the difficult detection characteristics of micro-small and slender defects,a feature extraction network supporting multi-scale feature extraction and fusion is designed.At the same time,based on this structure,it is proposed to use cross-channel and spatial dimensions to capture long-distance information between global features.While enriching feature information,the method improves the effectiveness and scale of feature information selection.In addition,in view of the large number of hyper-parameters contained in the complex feature extraction network,which is likely to cause the problem of slow algorithm training convergence speed,the isomorphic branch structure is introduced to further optimize the feature extraction network.(2)In view of the incompatibility between the original anchor frame generation size and the industrial panel defect samples,the hyper-parameter values are reset,and a multi-scale anchor frame generation network is proposed,which improves the prediction of extreme aspect ratio and small proportion defects The degree of match between the frame and the real frame.(3)This paper further studies the area pooling method,and the introduction of the area integration-based pooling method effectively avoids the lack of the original secondary quantization method that loses the feature information of the candidate area.Finally,through the introduction of various data increasing methods such as mosaic,the diversity of defective samples is effectively increased,and at the same time,the problem of over-fitting due to insufficient number of samples in the algorithm training process is avoided.Through the performance test on the industrial panel defect data set,the effectiveness and superiority of the defect detection algorithm proposed in this paper are verified.Experimental results show that the average accuracy of the optimized defect detection algorithm is 80.9%,which is 4.7%higher than the original algorithm.At the same time,small and medium-sized defects are detected,and the average accuracy rate has been improved to a certain extent.The simulation test was performed on the test set,and the recognition accuracy rate reached 96%,and from the test results,the defect recognition confidence was above 95%,which proved that the optimized defect detection algorithm can more effectively complete the surface defect detection of industrial panels Identification and classification tasks. |
PDF Full Text Request