Research On Surface Defect Detection Of Highly Reflective Metal Workpiece Based On Deep Learning

For the defect detection of the high reflective metal surface,compared with traditional manual visual inspection,the visual inspection performs well in many aspects,including high-precision quantitative detection,partial qualitative testing and target position detection.However,due to the high reflective characteristics of the metal workpiece's surface,there is more high light noise in the workpiece image,which is easy to cause machine error detection.Therefore,the defect detection accuracy of high reflective metal workpiece based on machine vision still needs to be improved.Aiming at the defect detection on the surface of the high reflective metal workpiece,this paper mainly studies on the following three aspects: image preprocessing,defect recognition of image,and defect classification based on depth learning.1.Firstly,the illumination unevenness of the surface image of highly reflective workpiece is corrected,and the algorithm based on two-dimensional gamma function is adopted to process the image.The RGB image is grayed and a variety of filtering algorithms are analyzed based on time domain and frequency domain.The research on denoising is mainly based on BM3 D filtering algorithm,and an improved BM3 D filtering algorithm combined with wavelet decomposition is proposed later.After processing the workpiece image through the above methods,the experimental results reveal that the illumination unevenness correction algorithm can better balance the illumination components in the image,and the improved denoising algorithm can effectively filter the specular noise points and enhances the texture information of ROI region.2.Fast detection of defects on the surface of the workpiece are conducted through edge extraction and feature analysis.In the process of edge extraction,the subpixel edge detection method is applied to the image defect.Several subpixel edge extraction methods based on moments are compared and analyzed,but the Zernike moment subpixel edge extraction algorithm is mainly studied,and a gradient edge model is proposed to improve the original algorithm.The gray level co-occurrence matrix is applied to analyze the defect features,and the threshold is set to determine whether the workpiece is good.The experimental results show that the method can not only detect efficiently,but also ensure the detection rate.3.The current mainstream classification network based on deep learning is studied with emphasis on analyzing the ResNet and DenseNet network models.On account of ResNet converge faster and fluctuate less in the training,the network model of ResNet50 is selected for defect classification detection.The network performance is optimized by the research of accelerate model training and transfer learning during training.The experimental results show that under the condition of hardware,the large batch size and low precision training can accelerate the network's training speed and convergence speed,meanwhile improve the classification accuracy.After the network model is fine-tuned through transfer learning,the convergence speed and final accuracy of the network model during training are improved accordingly.Based on machine vision detection and deep learning classification network,a fast defect detection method based on subpixel edge extraction is proposed in this paper,and ResNet50 network is used to classify workpiece defects.Compared with manual detection and traditional machine vision detection methods,the defect detection model of highly reflective metal workpiece performs well in anti-interference and detection accuracy,eventually it can achieve the classification accuracy of 98.3%.It is of great significance to promote the visual automatic detection of highly reflective metal workpieces in industry.