Research On Workpiece Surface Defect Detection Technology Based On Deep Learning

In the complex production process,it is difficult to avoid producing inferior workpieces with defects on the surface.The defect on the surface of the workpiece not only affects its appearance quality,but also affects its service life and stability,and at the same time affects the quality of the product obtained as a secondary process.This paper focuses on the two major problems of the deep learning-based workpiece surface defect detection algorithm from the perspectives of weakly supervised learning and supervised learning,and experimental verification is carried out on three workpieces: diode glass bulb,magnetic tile and hot-rolled strip steel.Aiming at the problem that the workpiece surface defect sample set does not contain border label information and the sample labeling cost is high,this paper implements a weakly supervised workpiece surface defect detection algorithm based on the improved NASNet.Firstly,in order to avoid over-fitting,the NASNet network is tailored;secondly,in order to improve the classification accuracy of the algorithm and thus the positioning accuracy,a bilinear structure is introduced;finally,the Grad-cam algorithm is combined to achieve defect location.Experimental results show that the improved algorithm improves the classification accuracy and reasoning speed,and at the same time can complete the surface defect location of the workpiece,and provides a feasible method for detecting the surface defect of the workpiece for the lack of border marking information.Aiming at the problem of low detection rate of difficult samples such as small targets in workpiece surface defect detection,this paper is based on improved SSD to realize workpiece surface defect detection.First,in order to reduce the model parameters and improve the model inference speed,the backbone network was replaced with Mobile Netv2;secondly,in order to improve the detection accuracy of small targets,a shallow feature fusion module was designed;finally,in order to solve the category imbalance and improve the detection accuracy of difficult samples,Introduce the Focal Loss loss function.The parameters of the improved algorithm model are only 7.23 M,the inference speed reaches 79.34 FPS,and the detection accuracy on diode glass bulbs,magnet tiles and hot-rolled strip steel workpieces reaches 96.94%,95.33% and 63.24%,respectively.The experimental results show that comprehensive model detection accuracy and complexity,the improved algorithm in this paper has better performance than other comparison algorithms.In addition,this paper implements a workpiece surface defect detection software system,which is convenient for visualizing the detection results.