Abstract: To address the issues of low recognition efficiency and accuracy in traditional weld defect detection methods when processing large-scale industrial data, a lightweight weld defect recognition algorithm based on convolutional neural networks was proposed. The fire module was introduced into the original MobileNetV3 to reduce parameter size, while the ECA (efficient channel attention) module was incorporated to enhance feature channel learning capability, thereby optimizing computational resource allocation and improving feature extraction performance. To validate the effectiveness of the proposed algorithm,comparative experiments were conducted with common classification models on a weld defect test dataset. The results demonstrate that, compared to other classification models, an average recognition accuracy of 98.50% is achieved by the proposed algorithm for common industrial defects such as dents, pores, and burrs, with the original algorithm being significantly outperformed, thanks to the combined effects of the fire module’s lightweight design and the ECA module’s feature enhancement. Moreover, both parameter size and floating-point operations are significantly reduced by the improved MobileNetV3 algorithm while high recognition accuracy is maintained, making it particularly suitable for deployment on industrial inspection devices with limited computational resources. A practical solution is thus provided for real-time quality inspection in the field of intelligent manufacturing.