Leveraging Generative Adversarial Networks and SE-UNet for High-Precision Tobacco Leaf Image Segmentation and Blend Uniformity Detection

作者：X. -S. Peng · 发表于：Zenodo (CERN European Organization for Nuclear Research) · 年份：2026 · DOI：10.5281/zenodo.20604426 · 研究领域：Advanced Neural Network Applications、Smart Agriculture and AI、AI in cancer detection

This repository contains the core experimental code associated with the manuscript entitled “Leveraging Generative Adversarial Networks and SE-UNet for High-Precision Tobacco Leaf Image Segmentation and Blend Uniformity Detection”. The deposited code supports the main computational workflow described in the manuscript, including tobacco shred image preprocessing, GAN-based image generation for data augmentation, semantic mask processing, SE-UNet model construction, model training, prediction, and evaluation. The proposed segmentation model is based on the U-Net framework with an integrated Squeeze-and-Excitation (SE) attention module, which is used for channel-wise feature recalibration in tobacco shred semantic segmentation. In addition, DCGAN, WGAN, and CycleGAN implementations are included as auxiliary image generation methods for expanding the training dataset. The purpose of releasing this code is to improve the reproducibility and transparency of the computational procedures used in the study. The repository includes the main scripts required to reproduce the model architecture and experimental pipeline. Due to data management restrictions associated with the cooperative enterprise, the complete raw tobacco shred image dataset is not publicly released in this repository. Representative sample data or annotation examples may be provided where permitted, and the full dataset may be available from the corresponding author upon reasonable request and with permission from th...