Abstract: In the process of automatic welding, welding point needs to be recognized with the help of laser. However, it suffers from the arc light and reflect light on the surface of some materials and the resulting accuracy of recognition cannot be guaranteed. In terms of this issue, the recognition network based on heatmap is proposed with combination of deconvolution and feature pyramid network. It extracts pyramid feature using residual convolutional neural network and generates key-point heatmap for each scale, which can tell the exact position of welding point. Compared with template matching and original feature pyramid network, such network performs better in the recognition of welding point with strong robustness and can work well in the context of various noise and complex interference.