Detection Of Coated Particles Based On Machine Vision

As the smallest unit of nuclear fuel particles,Coated fuel particle consists of UO₂core and the outside four layers with different density,thickness of different silicon carbide and pyrolytic carbon coating layer.Each particle’s diameter is only about 1 mm.In order to ensure the efficiency and safety of the future nuclear reactor operation,it is of great significance to accurately measure the thickness of coated layer of coated particles.Before measuring the thickness of coated fuel particles,it is necessary to identify and locate the measurable coated particles in the image.Therefore,it is very important to carry out research on the vision-based automatic localization and detection of coated fuel particles.The selected topic of the graduation project is derived from the scientific research project of the enterprise.Based on the characteristics of high efficiency,high objectivity and non-contact of machine vision,an imaging scheme of coated fuel particles is designed to image the metallographic plates.This paper focuses on the comparison of the detection results of a variety of target detection algorithms for the measurable coated sphere particles.In order to achieve better detection results,this paper also proposes an improved Faster R-CNN to realize the real-time detection of the measurable coated ball fuel particles.The main research works of this paper are as follows:（1）A traditional target detection method is designed according to the requirements of the coated ball particle detection project.Firstly,the initial image is preprocessed to separate the coated particles from the background;secondly,the center of the coated particles in the image is obtained by marking the connected region,and the candidate region is extracted based on the center of the circle;finally,the detection and location of the measurable particles are realized according to the gray histogram information of each candidate region.The experimental results show that the traditional target detection model can locate the fuel particles with high precision.（2）In view of the poor robustness and efficiency of the traditional target detection algorithm,a deep learning based detection model of coated sphere particles is proposed.Deep learning algorithm can obtain higher-level feature information through convolution neural network,so as to improve the robustness of detection model.Therefore,four classical deep learning target detection frameworks,R-CNN,Faster R-CNN,YOLO-V3 and SSD,are used in the experiment.The results show that the deep learning algorithm is superior to the traditional target detection algorithm in each performance evaluation index,and the Faster R-CNN algorithm in the four classical deep learning frameworks is more in line with the requirements of the coated ball particle detection project.（3）In order to improve the detection performance of Faster R-CNN algorithm,this paper also proposes three improved strategies according to the characteristics of the project.The transfer learning method can accelerate the speed of training model;the convolution neural network is also adjusted according to the super pixel industrial image;the size of candidate frame is also redesigned according to the size of the coated particle.Experimental results show that compared with the original Faster R-CNN algorithm,the improved detection algorithm improves mean IOU by 0.014 and detection efficiency by 30.2%.