Study on strip segmentation in complex scenes based on improved U-Net model

Abstract

Detecting PVC adhesive strips on the underside of automobiles presents significant challenges, including uneven illumination, pronounced reflections, irregular morphologies, and a high degree of similarity between target regions and surrounding pixels. Traditional techniques, such as threshold segmentation and edge detection, often fail to yield precise segmentation outcomes under these challenging conditions. To address these challenges, this paper introduces an enhanced U-Net model featuring three pivotal optimizations: Firstly, the U-Net backbone network is replaced with VGG16, which significantly enhances the model’s ability to extract multi-scale features, delivering exceptional performance in handling complex lighting conditions and irregularly shaped targets. Secondly, the Squeeze-and-Excitation (SE) channel attention mechanism is integrated, which empowers the model to dynamically emphasize critical feature channels, effectively suppresses background noise, and achieves higher segmentation precision. Finally, a composite loss function combining Dice Loss and Focal Loss is devised to effectively mitigate class imbalance and enhance the model’s capability in capturing fine edge details. Experimental results demonstrate that the enhanced U-Net model outperforms the traditional U-Net across multiple evaluation metrics, achieving a mean Intersection over Union (mIoU) improvement from 0.80 to 0.85. While maintaining computational efficiency, the model achieves highly precise and robust adhesive strip segmentation, showcasing significant potential for industrial applications, particularly in automated quality inspection within complex industrial environments.