Track foreign object image augmentation based on the proposed PLCA-pix2pixGAN method

The presence of foreign objects on railway tracks poses serious safety risks and may lead to accidents or service disruptions. However, existing detection systems based on deep learning are often constrained by small datasets, limited sample diversity, and low realism in synthesized training images. To address these issues, this paper proposes PLCA-pix2pixGAN (Perceptual Loss and Channel Attention Enhanced pix2pix GAN) to generate high-quality synthetic images for data augmentation. The method overlays object templates onto real-world track images to build a composite dataset and applies interpretable augmentation to simulate lighting and weather changes. To enhance fidelity, a channel attention mechanism enables region-aware reconstruction, and a multi-objective loss combines perceptual loss with adaptive weighting to balance pixel-level accuracy and semantic consistency. Experiments show the proposed method achieves an average SSIM of 0.9106 across object categories, demonstrating its effectiveness in generating realistic, structurally consistent images for safety-critical foreign object detection in railway systems. • Built a dataset combining real scenes and obstacle compositing with interpretable augmentation. • Developed a generator with channel attention to enhance key object feature learning. • Designed a multi-loss combining perceptual loss and weighting for better detail recovery.