Abstract
Leaf vein segmentation is a key technology supporting quality inspection and processing in the tobacco industry. However, factors such as wrinkles, occlusions, and complex secondary vein structures in acquired tobacco leaf images affect the accuracy of vein segmentation, making the reconstruction and restoration of veins after segmentation particularly important. Existing restoration methods suffer from inadequate accuracy in distinguishing pixels within fractured regions and coarse evaluation of local restoration, failing to meet the requirements for precise repair. To address this issue, this paper proposes an advanced leaf vein restoration algorithm named GDCA-Net. First, to tackle the difficulty in distinguishing between valid and invalid pixels inside and outside fractured areas, a multi-level generator network based on a U-Net structure is designed to expand the receptive field, capturing multi-scale contextual information and improving the quality and continuity of vein structure restoration. Second, to overcome the instability and lack of refinement in evaluating locally restored regions during adversarial training, a spectrally normalized Markov discriminator (SN-PatchGAN) is designed. Spectral normalization stabilizes the adversarial training process and refines the assessment of texture and structural consistency between restored regions and real images. Meanwhile, a multi-task composite objective loss function is constructed to enhance detail preservation and structural integrity. Experimental results show that GDCA-Net performs excellently in leaf vein breakage restoration, achieving a peak signal-to-noise ratio of 29.16 and a structural similarity index of 0.9962, outperforming other restoration methods and providing complete leaf vein data for accurate assessment of tobacco leaf stem content.