Purpose: Traditional inspection of railway turnout rutting machines, crucial for track switching safety, suffers from fragmented mechanical/electrical data, environmental limitations, and inadequate modeling of signal-state relationships. Existing intelligent algorithms lack visual characterization of mechanical deformation. This study aims to overcome these limitations and data scarcity to advance intelligent rail maintenance. Methods: We constructed the TurnoutSegmentation dataset with 2,360 high-definition, EN 50129-compliant trackside images. It features standardized pixel-level annotations of key components [Switch rail, Stock rail, Tie Rod (Switch Machine Rod), and Switch Machine] and is divided into training, validation, and test sets. Imaging incorporated dynamic light compensation for all-weather HDR, with pre-processing including grayscale equalization, ROI extraction, and geometric correction. Comparative experiments used the UNet architecture, evaluating segmentation performance. Results: The UNet++ model, optimized via deep supervision, achieved a rutting equipment segmentation accuracy of 0.8306 mIoU. This enables automated defect detection through precise part segmentation and establishes a baseline for cross-domain analyses like track condition assessment and predictive maintenance. Conclusion: This study breaks the bottleneck of data scarcity and algorithm generalization for turnout monitoring. The methodology promotes intelligent rail maintenance system development and is extendable to other industrial inspection scenarios requiring fine-grained analysis. The standardized dataset ensures reproducibility and supports vision-based infrastructure monitoring research. (Code: https://github.com/louzongzhi/Segment, Dataset: https: //doi.org/10.57760/sciencedb.22706).