| Object detection algorithm is a hot topic in the field of computer vision in recent years.With the development of deep learning,more and more target detection algorithms based on deep learning are gradually on the stage,which are widely used in production and life,and accelerate the process of social intelligence.Object detection is a challenging research topic because it contains many cutting-edge knowledge in many fields,such as pattern recognition,image processing,deep learning and so on.Considering that at present,the detection of foreign bodies in trains is mainly based on human visual inspection,which is labor-consuming and inefficient,this paper uses the improved algorithm of object detection based on YOLOv3 in this scene,which saves manpower and improves the detection efficiency,highlighting the application value of the theoretical research of object detection.In this paper,the object detection algorithm of YOLOv3,which is widely used and has good performance in recent years,is introduced.The backbone network and loss function of yolov3 are analyzed in detail and in-depth.And in view of the shortcomings of the corresponding improvements,and proposed the improved target detection algorithm Advance-YOLO,and in Pascal VOC 2007 data set experiments,to verify the effectiveness of the improvement.The main improvement points are as follows:(1)Using Dense block with high feature reuse rate and less parameters for training to build a new backbone network.Compared with the way of using residual module to build backbone network in the original YOLOv3,the method of this paper encourages the network to use the shallow,middle and deep feature information for comprehensive judgment and feature extraction,which significantly improves the ability of backbone network to extract image features,and enhances the ultimate target detection performance.The experimental results show that this single improvement can increase the m AP by about 2.1%.(2)we use the distance between bounding box and GIo U to redesign the bounding box loss in YOLOv3.It solves the problem that the loss value proportion of large and small targets caused by the loss of the original MSE bounding box is unbalanced.Furthermore,the detection performance of YOLOv3 for small targets is optimized.The experimental results show that this single improvement can increase the m AP by about1.05%.(3)In order to solve the problem of sample imbalance in the original YOLOv3,the cross entropy loss function of confidence in the original algorithm is replaced by the focal loss function.It improves the loss proportion of positive samples and hard samples,and enhances the learning ability of the network to the hard samples and positive samples.The experimental results show that this single improvement can increase the m AP by about 0.8%.Finally,combined with three improved Advance-YOLO object detection algorithm,the map reaches 78.15%.Compared with 74.85% of the original YOLOv3,the detection performance is improved by 3.3%,which is significantly improved.Finally,the improved Advance-YOLO algorithm is applied to the actual engineering scene: train foreign body detection.The problem of small sample data set is studied and solved in engineering application.In view of the complex field environment and the small number of foreign matter samples in the train,three strategies,data enhancement,K-means clustering prior box and transfer learning training,have greatly improved the training effect on the small sample data set and solved the problem of small sample data set.Finally,through the developed application program,based on Advance-YOLO train foreign body detection algorithm,to meet the actual needs of the railway site.The results of field installation and six months of practical test show that the accuracy of the algorithm developed in this paper is 63.5%,and the recall rate is 96.9%.It can effectively replace human to detect the foreign matters on the train body and meet the needs of users. |
PDF Full Text Request