<p>Fabric defect detection is a critical step in process control and quality assurance within the textile industry. To address the inefficiencies, inconsistent accuracy, and subjectivity of traditional manual inspection, this study proposes a lightweight defect detection approach based on deep learning. An improved ShuffleNetV2 architecture is developed, incorporating a dual-branch attention mechanism. The model enhances channel-wise representation via the ECA module and strengthens local spatial feature extraction through the LRSA module, thereby improving the detection of complex and subtle defects. Partial convolution is employed in place of standard convolution to reduce computational costs while retaining critical spatial information and maintaining compatibility with attention mechanisms. Experimental results demonstrate that the proposed model improves average precision by approximately 3.4% over the baseline while reducing the number of parameters by about 25%. Moreover, the model maintains high accuracy and real-time performance, meeting the practical requirements of industrial applications. The final improved model achieves effective control over size and computational complexity compared to conventional convolutional neural network architectures, offering a viable solution for real-time quality inspection in industrial settings.</p>