Abstract: In response to the issues of low efficiency and accuracy in traditional weld defect detection when processing large volumes of weld defect data, a lightweight weld defect recognition algorithm based on convolutional neural networks was proposed. Building on the original MobileNetV3 algorithm, the fire module was introduced to reduce the number of parameters in the algorithm. Additionally, the ECA (efficient channel attention) module was incorporated to enhance the learning of feature channels, enabling the algorithm to allocate computational resources more effectively and improve overall feature extraction capabilities. The improved algorithm was compared with common classification model algorithms on weld defect test data to validate its feasibility. The results show that, compared to the original algorithm and other classification model algorithms, the improved algorithm, with the combined effects of the fire module and the ECA module, achieves a further increase in weld defect recognition accuracy, with an average recognition accuracy of 98.53% for defects such as depressions, holes, and burrs. The improved MobileNetV3 algorithm not only maintains high recognition accuracy but also significantly reduces the number of parameters and floating-point operations, facilitating the lightweight deployment of the model.