A rapid packaging defect segmentation method based on enhanced UNet architecture

To overcome the excessive computational complexity of UNet networks in segmenting transverse seal opening defects in packaging, which limits practical industrial deployment, a lightweight LPGUNet segmentation network model was developed. The proposed architecture incorporates a multi-scale inverted residual feature extraction module (PCP), which combines partial convolutions and dilated convolutions, along with an embedded attention mechanism (CGBlock) to enhance segmentation accuracy. Ablation experiments validate the effectiveness of these architectural enhancements and the integrated attention mechanism. Performance evaluation shows that the model achieves an Intersection over Union (IoU) of 85.03% and a Pixel Accuracy (PA) of 91.61% in segmenting both background and defects. Comparative analysis with lightweight segmentation networks ENet and CGNet demonstrates that the proposed model offers superior overall performance, achieving higher computational efficiency and improved segmentation quality.