Recognizing Train Sand Pipe Based on the Improved PIDNet

The sand pipe is a key component of train sand blasting, and its height may change under the influence of locomotive and external factors, and the abnormal height will affect the sand blasting effect, and in serious cases, it may lead to wheel skidding, which may cause safety accidents. In this paper, an efficient algorithm based on improved PIDNet network is proposed for the task of semantic segmentation of train sand pipe and rail surface. Firstly, an efficient local attention mechanism module (ELA block) is introduced in the downsampling process of integral branch to reduce the loss of local information, so as to improve the segmentation accuracy of the details of the target region. Meanwhile, the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3 \times 3$</tex> convolution in the PAPPM module is replaced by the lightweight and efficient GSConv module, which enhances the multi-scale feature fusion capability and reduces the computational cost. Experiments show that the improved model significantly improves the accuracy metrics of PA and mIoU compared with the original PIDNet, achieving 96.85% PA and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$94.36 \% \text{mIoU}$</tex>, and the inference speed is 10.5 ms, which is better than the current mainstream methods.