Abstract: segmentation，high-precision component localization，and the necessity to accommodate products of the same category but differing models. Therefore，this paper proposes a vision localization algorithm for robotic arms based on a fusion network of CNN and Transformer. The proposed fusion network effectively utilizes global and local features to segment targets；The post-processing algorithm for localization calculates the grasping center and rotation angle based on the contour features of segmented images；The introduced edge feature enhancement module enhances the features of segmented edges；The loss function measures the deviation between the segmentation edge and the true value，as well as the deviation between the calculated center point and the true value. Experimental results on custom dataset demonstrate that the proposed method achieves a mean Intersection over Union of 0.987. The Mean Absolute Errors for grabbing center x，grabbing center y，and rotation angle A are 2.78 pixels，5.75 pixels，and 2.43°respectively， outperforming other state-of-the-art methods. Furthermore, comparative experimental results on the industrial component dataset T-LESS demons that the proposed method exhibits good applicability.