Lightweight railway panoramic segmentation based on channel shuffle and cross scale enhancement

A lightweight railway panoramic segmentation method based on channel shuffle and cross scale enhancement is proposed to address the issues of low accuracy and difficulty in achieving lightweight real-time segmentation in high-speed railway scene image panoramic segmentation. Firstly, based on the FasterNet network, a lightweight CS-FasterNet feature extraction network is proposed. The original FasterNet is a feature extraction network that reduces redundant calculations through partial convolution, aiming to improve computation speed and preserve high detection accuracy. However, some convolutions only apply filters to a small portion of the input channel, which can lead to insufficient feature extraction for the remaining channels. The aggregation of feature information was optimized through partial convolution and channel shuffling, and feature recombination techniques were used to reduce the computational complexity of the model, thereby achieving the goal of feature extraction in lightweight railway scenes. Realize lightweight feature extraction for panoramic segmentation in railway scenes. Secondly, a multi-scale feature interaction enhancement module was designed, which applies multi head attention mechanism to further extract pixel level semantic feature information from high-level semantic features of the image, which can expand the range of receptive fields, increase the long-range dependency relationship of feature maps, and help capture finer grained information. The cross scale feature fusion module adopts two paths of feature fusion: bottom-up and top-down. By fusing the feature maps of different scales output by the backbone network, it improves the utilization of scale features, comprehensively captures local details and global information, and which improves the quality of image feature extraction. Finally, the improved prediction fusion module integrates semantic results with instance results. In the panoramic segmentation task, the Soft NMS method is used to improve the accuracy of pixel classification. Soft NMS selects detection boxes by reducing the confidence of the intersection to union ratio greater than the detection box. The obtained intersection to union ratio value is multiplied by the original score using Gaussian exponent, and finally the true detection box is selected by weighting. Improve the accuracy of image segmentation by the network and obtain more refined panoramic segmentation output results. The experimental results show that the proposed lightweight model uses frame rate processing per second and computational complexity as evaluation indicators. The larger the value of frame rate processing per second, the faster the segmentation speed, and the smaller the value of computational complexity, the more favorable it is for lightweight segmentation. In the model, the evaluation indicators such as processing frame rate per second and computational complexity are superior to the comparison method. Compared with the UPSNet method, the processing frame rate per second of this method has increased by about 11.5frames/s, and the panoramic segmentation quality has improved by about 9.9%. It can achieve accuracy and real-time panoramic segmentation in different railway scenes.