Embroidery artefact image restoration technology based on improved DenseNet and GAN

Embroidery cultural relics often suffer from missing stitches and colour fading due to environmental and human factors. Traditional manual restoration is time-consuming (over 200 hours per piece) and prone to secondary damage. Existing deep learning methods struggle with structural distortion (>15%) and poor semantic consistency in complex textures like gold/silver threads and three-dimensional embroidery. To address these issues, we propose an advanced embroidery image restoration method. A DenseNetwork with channel and spatial attention achieves 94.25% accuracy, 93.47% recall, and 96.59% specificity for classification. Restoration uses an improved GAN with dilated convolutions, attention modules, a mask-guided discriminator, and joint loss. On datasets with 20-30% masking, the SSIM reached 0.971 (vs. 0.873 for traditional GANs). At 40-50% masking, the FID dropped to 17.33 (vs. 20.14). The model is efficient, requiring only 10.52G FLOPs, 5.58M parameters, and 0.62s per image. This method enables high-quality, efficient restoration of embroidery artefacts.