Research On Detection Method For Rail Fastener Defects Based On Deep Learning

As a low-cost and high-efficiency method that can replace manual inspection,the visionbased rail fastener defect detection method has great applicable value in railway transit safety.Due to the use of low-level visual features,the traditional methods have the characteristics of low accuracy,poor robustness and weak characterization in fastener location and defect classification tasks,it is not suitable for fastener defect detection task in complex environments.Compared with the traditional methods,deep-learning based methods have shown great performance in all aspects.However,most of these methods ignore the characteristics of the fastener and use the feature extractor with complex structures,thus they cannot achieve the detection speed that meets real-time requirements in the simple fastener location task.In addition,they cannot accurately locate the missing fastener regions.In the fastener defect classification task,most networks are structured by single-scale convolution kernels,and the extracted features are lack of discriminative contextual information.Moreover,due to the lack of defect fasteners,deep convolutional neural networks can easily lead to the over-fitting problem in a fastener training dataset with few samples.This paper studies the fastener defect detection method based on deep learning.To solve the problems in the previous deep-learning based methods,this paper proposes a fastener location and a fastener defect classification algorithms based on deep convolutional neural networks,and evaluate the performance of both algorithms on the non-public dataset provided by the Academy of Railway Sciences.The contributions of this paper can be summarized as follows:(1)A fastener location algorithm based on feature fusion attention mechanism is proposed.This algorithm uses a location framework based on center points detection,which can reduce post-processing time.Moreover,a lightweight hourglass-shaped network is designed as the feature extractor,and the skip connection and attention mechanism are introduced to design a feature fusion attention module which is used to improve the feature extractor.This lightweight network can extract features with rich semantic information and high resolution,and contributes to fastener location with high precision and speed.In addition,a post-processing algorithm based on prior knowledge is proposed to locate the missing fastener.Compared with other deep-learning based methods,the algorithm proposed in this paper achieves an optimal balance between location precision and speed in the fastener location task.(2)A fastener defect classification algorithm based on an efficient multi-scale feature selection mechanism is proposed.This algorithm first proposes an improved efficient multiscale feature selection module,which compresses the parameters of the module by an efficient structure design so that it can extract multi-scale features with rich contextual information and alleviate the model over-fitting problem.Secondly,a learnable feature affine transform layer is proposed to further improve the generalization of the model by augmenting the feature diversity.Furthermore,a fastener defect classification network is proposed based on the efficient multi-scale feature selection module and the feature transform layer.Finally,a Joint Loss is proposed to optimize the intra-class and inter-class distance so that the network can extract more discriminative features.Experiments have shown that efficient multi-scale feature selection module,feature transform layer and Joint Loss can effectively improve the feature expression ability of the network and solve the over-fitting problem.Compared with other methods,the fastener defect classification algorithm proposed in this paper has achieved the highest accuracy,and has better generalization.