In railway turnout sliding bed plate damage detection, manual inspection is inefficient with high missed detection rates, while existing target detection algorithms struggle with small defects under complex lighting and scarce samples, leading to poor generalization. This study proposes an improved YOLOv8-based detection method. First, a defect image enhancement model addresses uneven training sample distribution. Second, an improved YOLOv8 combining deformable attention and an enhanced loss function captures tiny defect features. Experiments are conducted on a mixed dataset containing real-world railway-site images and WGAN-GP-synthesized defect samples, totaling 8200 images covering four typical sliding-bed-plate defect categories under diverse lighting conditions. Results show stepwise improvement across all defect types: shrapnel breakage (small target) improved most (precision +3.2%, recall +3.6%), while excessive wear and bottom plate fracture (large target) improved about 2.9–3.7%, verifying the optimized loss function's effect on bounding box regression. Mean precision improved +3.2% and recall +3.4%. This offers a feasible technical solution for intelligent railway turnout sliding bed plate damage detection, supporting the transformation toward "intelligent monitoring and active prevention.".