Research On Railway Tool Image Classification And Detection Method Based On Deep Learning

The railway system is the main method of material transportation and personnel transportation,and railway tools are the key to the maintenance and repair of the railway system.The management of railway tools is im portant to improve work efficiency.The primary task that needs to be solved is how to accurately Identify the tools needed.At the same time,the development of deep learning technology has achieved great success in the field of computer vision,and the research on image recognition is also in full swing.Under this background,this paper studies the railway tool image classification method based on deep convolutional network and the railway tool detection method for natural scenes.The main contribution of this paper is as follows:(1)Aiming at the problem of lack of existing railway tool image data sets,this work crawls railway tool images from Google Gallery according to GB tool standards,and constructs the Railway Tool-10 railway tool image data set after preprocessing,which contains 7975 railway tool images in 10 categories.At the same time,in view of the lack of railway tool image detection data set in natural scenes,this paper combines the Railway Tool-10 data set and the COCO data set to construct the Railway Tool-Detection railway tool detection data set through the data synthesis method.The data set contains10 categories,a total of 10,000 images of railway tools in natural scenes;(2)As the methods which classify railway tool images are not so much in the current image classification research,this article starts from the deep learning classification algorithm and aims to alleviate the problem that the loss of spatial geometric information of the deep features in the current Res Net classification model and the feature channel is not considered in the feature extraction process.For the issue of dimensional relations,SARNet(Skip Attention Res Net)is proposed.SARNet introduces spatial attention sub-module and channel attention sub-module in Res Net,learns the channel weights and spatial weights of features,and further uses skip method to fuse the deep and shallow features of the model to strengthen the spatial geometric information of deep features.In this way,the performance of Res Net feature extraction has been improved.Then,a comparative experiment has conducted based on the Railway Tool-10 dataset.The experimental results have shown that compared with the original Res Net,the classification accuracy of SARNet is increased by 7.2%.Finally,the interpretability analysis of the SARNet model is based on the Grad-CAM method.Grad-CAM can show which areas of the image that the model focuses on in the form of heat maps.The analysis results show that SARNet can accurately locate the key visual areas of the image;(3)There is rare natural scene-oriented railway tool detection method in the current target detection research.Starting from the deep l earning target detection algorithm,this paper proposes a multi-task structure for the problem of small targets and sparse features of railway tools in natural scenes.Under the Multi-task Railway tools Inspection Network(MRIN),the model is divided into two parts: 1)Multi-scale detail feature extractor,this module uses Res Net101 as the backbone network and uses an attention mechanism that adds the FPN structure to extract image features,which solves the problem of small targets and sparse features of railway tools in natural scenes;2)Multi-task railway tool prediction module,which is mainly based on the prediction module of Faster R-CNN and extracts features from MDFE to locate and identify the target area.Finally,the experimental verification is carried out on the Railway Tool-Detection data set.The experimental results show that the m AP value of MRIN on the Railway Tool-Detection data set reach 0.977.