Design And Implementation Of Screen Surface Defect Detection System Based On Improved Faster R-CNN Algorithm

In the information society,the screen has become the main source of information for the public,and the increasing popularity of mobile phones,computers,etc.has also brought the production and innovation of the screen to a new peak.People hope to obtain a more stable and higher resolution display,but in the actual manufacturing process,it is easy to produce defective products,and the quality of the screen during the production process is a very important step.How to detect screen defects efficiently and accurately is an urgent problem to be solved in the production process.At present,the quality detection methods of industrial screens generally have the disadvantages of time-consuming and low accuracy.How to improve the detection capabilities of screen defective products is a current research hotspot.Based on the research of the existing screen defect detection system,this thesis designs a defect detection scheme that combines traditional image processing algorithms and target detection technology.Collected a batch of screen defect data sets in real industrial scenes,divided them into four defect categories,and carried out labeling and augmentation of the data sets.The improvement work is based on the Faster R-CNN algorithm.In the case that the original algorithm could not converge,the data set was analyzed,and it was determined that the image size was too large and the target was too small.The segmentation module with overlapping sub-regions was proposed.Through the analysis of classification tasks and detection tasks,a new feature extraction network is proposed to increase the size of the feature map;the network receptive field is increased by adding an expanded convolution structure;multi-layer feature fusionimproves the network feature extraction.After these changes,the algorithm’s detection ability on the screen defect data set is effectively improved.The TensorRT technology is used to quantize and compress the trained model,which reduces the model size and running time.The screen defect detection scheme proposed in this thesis reaches 0.91 on the industrial real test set,which effectively improves the detection rate of defective products.Through horizontal comparison experiments with common detection methods,it can be seen that the screen defect detection scheme proposed in this thesis has obvious advantages in speed and detection accuracy.