| With our country advance on a path to a a powerful transportation country,due to the improvement of the overall technical performance of the train,the quality requirements of accessories are also further improved.As the key parts of the train,bolster and side frame castings need to be more strictly inspected and controlled.However,due to the inherent production shortcomings of casting,bubble and loose defects often appear in the process of casting production,which bring potential safety risks to train operation.Therefore,in the process of casting production,it is necessary to strictly check and control the defects of bolster and side frame castings.In this thesis,the defect intelligent detection and location algorithm is studied.According to the fact that there are a lot of small defect targets in the defect data set and there are many missing labels and wrong labels in the original data set,this paper proposes a solution of defect detection and improves the algorithm by combining convolution neural network and traditional image method.Automatic defect detection and recognition by computer vision technology aims to reduce the false detection rate and missed detection rate of defects and reduce the work intensity of workers on the assembly line.The main contents of this paper are as follows:Compared with the general object detection task,the defect target area in the casting defect detection task has the characteristics of small area and the common object detection algorithm often can not detect small defects.To solve this problem,through the following ideas to improve the accuracy of defect detection:1.Using image segmentation algorithm to retain the small defect target information,all image blocks are predicted separately and the results are summarized to the whole image.2.Designing feature pyramid network in detail for improving the expression ability of feature map detail information.3.Using differential evolution algorithm to adaptive search the anchor,so that the anchor can cover more small objects.4.Using small object replication data enhancement method to improve the frequency of small objects and the contribution of small objects to the loss function.This thesis conducts training and testing experiments on the collected data sets,and the results show that the above methods are effective to improve defect detection rate.In the original data annotation,there are many missing labels and wrong labels.Low quality annotation will optimize the network in the wrong direction,which will inevitably affect the performance of the network.This thesis proposes a data noise cleaning method based on active learning,which embeds the above object detection network into the active learning framework.In the running process,the network selects a certain number of images through the query strategy based on confidence and the query strategy based on annotation,and then puts them into the training set to train.At a lower cost of labeling,by cleaning the label noise and improving the quality of the data set,the experiment shows that this method can further improve the performance of the network. |
