Research On Detection Technology Of Workpiece Surface Defects In Large Field Of View Based On Deep Learning

In this era of intelligent information,intelligent manufacturing will be one of the indispensable directions for the development of manufacturing.In intelligent manufacturing,product inspection is an important part of it.With the rapid development of modern manufacturing,under the requirements of product industrial accuracy,the detection of surface defects of large-field products has been widely promoted.Based on this,this paper mainly conducts research from two aspects of large-field image data acquisition and defect detection.Aiming at the acquisition of large-field image data,this paper designs an optimal path algorithm for horizontal movement of camera shooting,so as to quickly and efficiently acquire image data and prepare data for experimental research.The design of the algorithm is completed under the thought guideline of "crude first,then refined,and search the reason by the result".First,roughly take the overlap rate of the two pictures as10% as the camera movement step to take pictures,and then set the overlap rate in increments of 10% and take pictures,and then perform simulation experiments based on the collected data.According to the relevant index values of the experiment,an approximate range of overlapping rate that can meet the stable image stitching is given.Then iteratively optimizes according to the relationship between the surface size of the workpiece to be photographed and the area covered by the field of view of a single shot,Refine the overlap rate between adjacent shot images,and then give the camera moving step.Through further analysis of the complexity and convergence of the algorithm,it can be seen that the algorithm complexity is O(n),which quickly converges.Combining observation and analysis of the image stitching experiment results,it can be seen that the algorithm designed in this paper has a certain degree of reference value in the direction of acquiring large-scale image data to be stitched.Aimed at the surface defect detection of workpieces,inspired by model migration learning and migration application ideas,under the guidance of model parameter migration application ideas,this paper uses sliding window cutting technology to establish a defect detection model that can be used to detect large-field images.The model is mainly composed of three parts,that is,first build a Darknet-53 deep learningnetwork architecture,establish a defect detection model based on Yolov3 algorithm,and then modify the fully connected layer of the model based on the sliding window cutting idea to construct a large-view workpiece surface Defect detection model.According to the experimental conditions,firstly,we train the defect detection model based on the Yolov3 algorithm.Then using the good training model tests the experimental data set.The average accuracy of the model performance evaluation index(m AP)and the positive sample recall rate(Recall)are both up to 90%,90.81% and 98.99%respectively.At this time,migrate the relevant parameters of the model and applied to our new model,and a defect detection experiment was carried out.The experimental results show that the original model has a large degree of missed detection and detectable defects in the defect detection in the large-field image The accuracy rate given has also decreased,and comparing with our new model,the missed inspection situation is improved and the accuracy of the same defect inspection is improved.In addition,combined with the practical value analysis of the model,it gives the floor thinking of applying the model to production.