Feature cross fusion method for foreign object images segmentation on high-speed train bogie frames

Aiming at the problems of redundant segmentation and target edge detail loss in the foreign object images of bogies of EMU, an optimized segmentation model called FCFNet is proposed. A multi-branch structure of ResNeSt is used to construct backbone feature extraction network, to improve the differentiation and diversity of irregular foreign object images. In the decoder, the convolutional block attention module mechanism is used to enhance the deep features to obtain the effective space and channel features. Feature cross fusion is performed on different layers of features to improve the problem of missing edges of foreign object images caused by insufficient multi-layer feature fusion. The focal loss function is used to adjust the weights of difficult and simple classification samples, so that the model pays more attention to the pixels that are difficult to classify, and improves the segmentation accuracy of foreign object images. Experiments show that the mean pixel accuracy and mean intersection over union of the proposed method in the task of bogie foreign body images segmentation are 88.79% and 86.92%, respectively, and the inference time of a single image is 2.8ms, which meets the requirements of engineering detection accuracy and real-time.