Research On The Detection Of Key Components Of Train Bottom Based On Deep Learning

The high-speed train is usually an organism composed of multiple components.The key components of the train are the important parts connecting all parts of the train.The detection,identification and reliability evaluation of the key components can provide a considerable guarantee for the safety of the train.The environment of train bottom components is complex and traditional detection methods detect small targets badly,and are easily affected by illumination,water stains and other factors,so the robustness is poor.Introducing the deep learning object detection algorithm into the key component detection of train bottom task,designing a deep learning algorithm suitable for the key component of train bottom detection task,and exploring how to improve the algorithm performance when the available annotation data is limited is the main work of this article.Firstly,a key component of train bottom detection algorithm based on Retina Net is studied.On the basis of Retina Net,the loss value of key component of train bottom dataset in each detection layer is visualized,the P6 and P7 detection layers with smaller loss value are removed,and the network structure and parameters are simplified to make it more suitable for key component dataset.The object size distribution in the dataset of key components is analyzed,and the RFB(Receptive Field Block)module is introduced after shallow feature P3 to improve the small target detection rate.Then,the PAN(Pixel Aggregation Net)module is introduced to make the feature fusion method better and higher positioning accuracy.Secondly,a key component of train bottom detection algorithm combining semisupervised learning and model ensemble is studied to improve the performance of the model when the number of annotations is small.Semi-supervised learning can improve the performance of single model by using unlabeled data in the way of pseudo label,and the improvement of performance depends on the quality of pseudo label.Multi model ensemble can improve the performance of model,but increase the reasoning time of algorithm.Combining the two methods,it is possible to generate more accurate pseudo labels through multiple model integration for semi supervised training,which is helpful to improve the performance of single model.Finally,when the labeled data available for training is sufficient,the proposed improved Retina Net algorithm has better performance on the public dataset Pascal VOC than Retina Net,and the m AP evaluation index is better than other networks in the key component detection task.Experimental results on public dataset and key component tasks show that the proposed model ensemble semi-supervised method can improve the performance of single model when the labeled data is limited,and is superior to the semi supervised object detection algorithm STAC.