The Research On Surface Defect Detection Method Of Industrial Products Based On Deep Learning

The surface quality inspection of industrial products is an indispensable link in manufacturing,and it can not meet the requirements of industrial automation production by only relying on manual quality inspection.With the development of image processing technology,vision detection schemes based on machine vision are widely used.Traditional quality detection methods based on machine vision excessively rely on manual features,resulting in weak model generalization.Meanwhile,due to small defects and similar backgrounds,the location of surface defects in industrial images cannot be located.With the rapid development of computer vision,vision schemes based on deep learning have been making breakthroughs in various visual tasks.However,due to the few defect samples,high labeling cost,large computation,and difficult model deployment,the application of deep learnbased detection methods in the industrial field are slow,and it is difficult to achieve large-scale application in actual production.In this paper,based on the analysis of existing target detection and segmentation methods,combined with the task requirements of industrial defect detection,the design of an industrial product surface defect detection algorithm based on deep learning.The main research contents are as follows:(1)Defect detection model based on light-weight multi-scale information fusion architecture.Firstly,a fusion structure based on segmentation network and classification network was designed.The segmentation network was used to locate defects at the pixel level.In order to capture small feature details in the segmentation network,pooling was used instead of large convolution to realize down-sampling.Binary image classification was performed using classification network.Multi-layer convolution kernel was used for down-sampling in the classification network to enhance the ability of the network to capture information,and the classification network was built on the segmentation network,and the residual short connection was made to the feature graph of the segmentation network to reduce the number of parameters.The global pooling layer was used to follow up the output of segmentation network and classification network to solve the problem of mismatch between the output dimensions of segmentation network and classification network.Aiming at the problem that streamline structure cannot accurately extract input data information,a lightweight structure based on multi-scale information fusion was designed to enhance the expression ability of feature graph,further reduce the number of parameters and reduce the cost of model deployment.Finally,experiments are carried out on three industrial datasets.The experimental results show that the detection accuracy of the model was higher than 99%on different datasets,and at the same time,it has a low false detection rate,which can meet the needs of rapid and accurate detection in the industrial field.(2)Light-weight defect detection model based on end-to-end learning.Aiming at the disadvantages of the cumbersome training process and slow learning of the fusion model,an end-to-end training model was designed,and the segmentation network and classification network were studied simultaneously to accelerate the network learning process.Pixel-level annotations were processed to reduce the model’s dependence on labels.The gradient flow from the classification network to the segmentation network was adjusted to effectively overcome the position uncertainty of rough labels and reduce the risk of unstable features damaging learning.The sampling method was improved according to the frequency of normal samples to improve the detection ability of the model.Finally,the model was evaluated,and the experimental results show that the model can achieve excellent performance with only a small number of samplers.(3)Surface defect detection model based on hybrid supervised learning.Aiming at the problems of few abnormal samples and high labeling costs in the practical application of the model,a surface defect detection model based on hybrid supervised learning was designed to explore the influence of training data with different mixing ratios on network performance.Self-attention module was used to guide feature generation to improve the accuracy of defect detection.Weighted segmentation loss was designed so that the model could pay more attention to the location with a high probability of defects.At the same time,experimental research was carried out on self-made rail datasets.Experimental results show that this method is superior to other relevant methods in fully supervised learning.Meanwhile,it can be found that the hybrid supervised model with a small amount of pixel-level labels in image-level labels has the same performance as the fully supervised model while reducing the labeling cost,which meets the requirements of high-precision detection of industrial surface defects.