In the industrial environment of laser cutting, there are a large number of complex backgrounds and other interfering elements such as equipment. The current combination of category learning methods lacks a separate target alignment process for the highlighted areas in the abstract target graphics, resulting in weak correlation between target features and inaccurate annotation results. Propose a visual image object annotation method for laser cutting robots based on improved YOLOv5s. By utilizing input terminals, pooling layers, and shared fully connected layers, an improved YOLOv5s model is built. This network uses max pooling and average pooling to generate two laser cutting robot visual images. Based on the channel dimension, the image features are connected to achieve rough target localization in the laser visual image, and target feature alignment is achieved by combining modulation factors and target detection losses. After aligning the target features, the key highlighted area frames of the laser cutting robot's visual image are determined. By implementing semi supervised training on the historical annotated images of the laser cutting robot, the spatial region associations of the images are determined, and laser visual image target annotation is performed based on the region associations. The experimental results show that the proposed method for laser cutting robot visual image target annotation has high intersection to union ratio, high accuracy, fast speed, and strong robustness.