<p>Semantic segmentation models have achieved significant success in fields such as medical image segmentation, while their large sizes and slow inference speeds restrict their application in the high-speed melt-spinning process of yarn production. To address this challenge and meet the demand for rapid and accurate online inspection of yarn quality in the textile industry, we introduce MiniUNet, a novel ultra-lightweight and efficient segmentation model specifically designed for deployment on mobile devices. In the encoder, we have implemented a new convolution block that uses symmetric stripe convolution with shared parameters to extract features without axial deviation while significantly reducing computational costs. In the bottleneck, we have constructed an enhanced attention mechanism and active capture technology of row-column structure information to replace high-parameter traditional convolutions. Additionally, a lightweight decoder is used to generate segmented images, achieving a balance between efficiency and performance. We have rigorously validated the effectiveness of MiniUNet through extensive testing on a real-world yarn dataset and an open-access dataset, International Skin Imaging Collaboration 2018. Compared to the benchmark UNet model, MiniUNet achieves a 99.90% reduction in parameters, from 31.03 million to 34.45 thousand on the yarn datasets and to 31.00K on an open-access dataset. Concurrently, it makes a 99.85% decrease in computational complexity, from 215.76 GFLOPs to 324.21 MFLOPs on the yarn dataset and from 46.02 GFLOPs to 68.61 MFLOPs on the open-access dataset. Moreover, it demonstrates a 3.42 times faster inference speed on the low-performance CPUs of industrial computers and 76.67 times on the Jetson Nano. The results highlight MiniUNet’s potential to advance yarn quality inspection in the textile industry but also its prospects in other industry high-speed real-time detection such as quartz and bearing surface defect detection.</p>