Railway Fastener Pixel-Level Detection Based on Dual-Stream Encoder Network With Mask Guidance

Fastener pixel-level detection can provide a reliable basis for the assessment of fastener defects and requires only normal samples. Deep learning technology has been widely used for fastener detection due to its powerful ability in feature extraction and self-learning. However, the performance of many existing deep learning-based methods still requires further improvement as they fall short in producing the accurate pixel-level detection, especially for the fasteners in complicated backgrounds. To tackle this challenge, a novel dual-stream detection network (DSD-Net) based on encoder-decoder architecture is proposed for fastener pixel-level detection. In encoding stage, the enhanced and emphasized features of fastener foreground can be obtained by the dual-stream (i.e., raw image stream and mask image stream) encoder embedding designed feature enhancement module and cascade residual pooling module. In decoding stage, the decoder aggregates the features from dual-stream encoder by the feature enhancement module with skip-connection to improve the final fastener pixel-level detection results. Numerous experiments on the constructed dataset demonstrate that DSD-Net achieves more remarkable detection performance (Precision of 96.54%, Recall of 97.62%, Accuracy of 96.46% and IoU of 94.32%) for fasteners against other state-of-the-arts.