GIRH-Unet: Improved Residual Tobacco Segmentation Algorithm Based on GhostNetV3-Unet

Cigarette production presents an industrial environment with significant light and shadow interference. The residual tobacco contains fine particles that exhibit random distribution and varying edge morphologies. These factors contribute to the reduced accuracy and robustness of visual detection technologies based on segmentation algorithms within tobacco intelligent production systems, highlighting the need for a targeted segmentation model. In this study, we propose an enhanced lightweight segmentation model, GhostNetV3-Unet, designed explicitly for residual tobacco segmentation. Our approach utilizes an improved GhostNetV3 to bolster feature extraction capabilities. Additionally, we optimize the residual modules and activation functions to enhance the extraction of detailed features and the deep transmission of information. Furthermore, we enhance the semantic representation of fine residue features amidst complex backgrounds by incorporating a hybrid self-attention mechanism that integrates spatial information and channels attention within the decoding phase. We employ Focal Loss combined with Intersection over Union (IoU) loss to address the class imbalance challenge between positive and negative pixel classes, thereby improving the segmentation performance in regions with morphological and size variations. Moreover, we designed an image enhancement method based on Poisson fusion to mitigate the difficulties associated with sample labeling. This method employs a combination of randomly cropped, combined sample images and Poisson-fused background images to enhance edge contour clarity while effectively expanding the tobacco dataset. The experimental results demonstrate that the constructed artificial data samples can effectively substitute the original data samples, significantly reducing the labeling costs associated with sample preparation. On the tobacco dataset, the interaction rate of the proposed GhostNetV3-Unet model reaches 94.63%, with a checking rate of 97.91%, while maintaining a model parameter count of 17.88 million. Additionally, on the bird and light-and-shadow datasets, the model achieves interaction rates of 92.03% and 92.86% and checking rates of 95.22% and 96.22%, respectively. Practical application in the production environment confirms that the proposed method successfully segments fine residual tobacco in complex backgrounds and meets the detection requirements relevant to the production site.