Abstract: In view of challenges such as minute defects and strong background interference in robotic welding, an intelligent defect detection algorithm fusing multi-scale edges and hybrid attention was proposed for the robotic welding information collection platform. First, a multi-scale edge enhancement module was designed to break through the limitations of the convolutional receptive field, strengthening the geometric contour perception of multi-scale irregular defects such as porosity and weld nodules. Second, a local-global channel attention mechanism was introduced to integrate local acuity with global contextual vision, dynamically filtering out background noise such as strong reflections and complex metal textures to acquire features with a high signal-to-noise ratio. Finally, a lightweight shared convolutional detection head was constructed, and by adopting group normalization and weight-sharing strategies, computational redundancy was substantially reduced while eliminating batch-size dependency. Tests indicate that compared with the baseline model, the number of parameters of the proposed algorithm is reduced by 6.6%; the mean average precision increases by 3.9% to reach 63.7%, and the inference speed reaches 135.5 frames per second. Heatmap visualization analysis confirms that the high-response areas of the network accurately converge on the faint targets and their geometric edges, successfully suppressing invalid background activations from the source. The proposed method achieves an excellent balance among precision, timeliness, and complexity, fulfilling the demand for highly reliable recognition.