Study On Casting Defect Detection Based On Deep Learning

As an indispensable part of modern industrial automation,castings are becoming more and more important in improving industrial production efficiency and realizing intelligent manufacturing.Casting equipment is widely used in many industrial fields,and the quality of castings has naturally become the top priority in industrial production.It takes a lot of manpower and material resources to detect the quality of castings,so as to ensure that the castings are in good working condition and reduce the risk of accidents.Therefore,there is an urgent need for an economical,efficient,accurate and intelligent casting defect detect method to ensure the efficient use of castings.At present,most of the casting defects detection still rely on manual detection,which costs a lot of manpower and time.At the same time,as the working time increases,the speed and accuracy of manual detection will be affected.In addition,with the advancement of the intelligent manufacturing process,higher requirements are put forward for the degree of intelligence of the casting assembly line,and the quality of the casting is also more and more stringent.Obviously,traditional manual detection methods can no longer meet the needs of automated assembly line production,and have become a bottleneck restricting its development.Manual participation hinders the realization of intelligent automated assembly lines.With the development of computer science and deep learning technology,the use of neural networks to detect defects in X-ray images of castings is an effective way to solve this dilemma.Therefore,in order to realize a truly intelligent automated production line and reduce the tedious manual labor in the production process,the research on the automatic detection algorithm of casting defects based on computer image and graphics technology is particularly important.Therefore,this master's degree thesis focuses on designing an algorithm framework that realizes casting defect detection(semantic segmentation)and defect location at the same time.Committed to realizing economical and efficient casting defect segmentation methods;Committed to exploring low computational cost and convenient defect location methods;Committed to real-time,accurate and integrated algorithm of segmentation and positioning.The main research contents of this paper include:1.Three training strategies are introduced into the single-stage target detection algorithm FCOS to verify the influence of various strategies on the network.And compare with the twostage network Faster R-CNN.Three different strategies are introduced to improve the performance of the algorithm for classification tasks without increasing the computational cost of the algorithm.Experimental results show that not all strategies can improve network performance.The experimental results also show that the accuracy of the two-stage target detection network Faster R-CNN in completing the casting defect detection task is higher than that of the single-stage target detection network FCOS,but the speed of the FCOS algorithm is significantly ahead of Faster R-CNN.2.In order to further improve the accuracy and speed of defect location,we found that it is necessary to improve network performance and reduce the amount of network parameters.(1)We manually annotated a dataset containing X-ray image defects of castings.The pixel-level instance and semantic labels of the dataset are given.In the future,we will choose the opportunity to disclose this dataset.(2)We propose an improved single-stage instance segmentation network beacon detection(improved YOLACT)instead of two-stage network.By adding feature fusion and supplement modules to the network,the feature utilization rate can be improved,and the network detection performance can be improved.And the algorithm was verified on the manually labeled dataset.The improved network has improved detection capabilities compared to the original basic network.