Research On Industrial Product Defect Detection Algorithm Based On One-stage Strategy

The symbol of national development is industry,and industry is the leading force in the development of the national economy.An important indicator to measure industrial development is production efficiency.One of the effective ways to improve production efficiency is to solve the bottleneck process.A common bottleneck problem in industrial production is the detection of industrial product defects.The algorithms currently used in the actual production environment for industrial product defect detection include color difference formulas,template matching and deep learning target detection algorithms.Among them,the color difference formula and template matching method have high environmental requirements,low robustness,and high threshold for use,which are not suitable for industrial product defect detection in actual production environments.The deep learning target detection algorithm is divided into one-stage algorithm and two-stage algorithm,with low threshold for use and strong robustness.The two-stage algorithm has high detection accuracy,but the detection speed is slow.The detection speed of the one-stage algorithm is much faster than the two-stage algorithm,but the detection accuracy is slightly lower than the two-stage algorithm.The two types of deep learning target detection algorithms cannot achieve a balance between detection accuracy and detection speed,and it is imperative to study fast and accurate target detection algorithms.Combining the objective needs of industrial product defect detection and the characteristics of the actual production environment,this article focuses on the classic one-stage model SSD’s network decay problem and the problem of too large model and slow detection speed.YOLOv3 feature output layer is less,and the problem is large.For the problem of poor target detection ability,a one-stage target detection model based on residual network and attention mechanism is proposed,which improves the speed and accuracy of model detection;for SSD and YOLOv3,due to insufficient semantic information in the shallow feature layer,For the problem of poor detection of small targets in target detection,a method based on deconvolution and feature fusion is proposed to improve the accuracy of the model’s detection of small targets;some commonly used loss functions are not scale-invariant,Some are not friendly to small targets,etc.,a loss function based on LWR is proposed.This loss function has scale invariance,is easy to train,and is friendly to small targets;for SSD,it is impossible to flexibly generate appropriate default bounding according to the data set.The problem of boxes’ size value,and the calculation of default bounding boxes’ size value by YOLOv3 is susceptible to outliers,and does not consider the size of the receptive field of the model feature output layer.This article proposes an improved default bounding boxes’ size value calculation algorithm combines the size distribution of the target in the data set and the characteristics of the model itself,and generates a set of suitable default bounding boxes’ size value for each data set to improve the detection accuracy of the model.Through the above work,the FA-SSD has reached a balance between speed and accuracy,so that the FA-SSD model has a faster detection speed and a higher detection accuracy.This paper chooses two public data sets Pascal VOC 2007 data set and DIOR data set in the field of target detection,and two public small and medium target data sets on the surface defects of industrial products: the surface defect data of hot-rolled steel strip released by Northeastern University.Set NEU-DET data set and German textile texture data set TILDA data set,a total of 4 data sets,in the same experimental environment,compare the FA-SSD model with SSD,YOLOv3 and DSSD to test the effectiveness of the FA-SSD model authenticating.Experimental results show that FA-SSD effectively improves the detection accuracy of small targets,achieves a balance between detection speed and detection accuracy,and performs well on both the target detection general data set and the industrial product defect detection data set.