Visual detection plays a vital role by enabling machines to learn from complex data and make informed decisions. In the context of “mixed line production—common line packaging” in the automobile wheel manufacturing industry, accurate wheel recognition is crucial for automated gripping, shelving, and subassembly, which also provides essential feed-forward information for various processing steps. However, within the same category of wheel hubs, there are often significant differences in painting techniques, which poses a challenge for the construction of a complete dataset while also leading to substantial intraclass variance, which seriously affects the accuracy of the recognition. Furthermore, wheel recognition requires not only accurate classification of known wheels but also timely detection of unknown wheel mixed, which leads to the inapplicability of the traditional pattern recognition methods based on the closed-set assumption. To address the above challenges, the article proposes a two-stage solution combining data generation and metric learning. Specifically, for the problem of missing data from some painting processes, a CycleGAN model based on edge consistency and structural similarity (ECSS-CGAN) is proposed, which realizes the generation of wheel images with missing painting processes to construct a high-quality pseudo-complete dataset. Then, a Gaussian kernel function-based deep metric learning network (GKDML-Net) is proposed to address the problem of substantial intraclass variance of the same wheel due to different painting processes. Moreover, to increase the robustness of the proposed method with the pseudo-complete dataset, the Gaussian kernel loss and center loss are incorporated to extract highly compact discriminative features. Finally, the classification of known wheels and the detection of unknown wheels mixed in the actual production line are achieved simultaneously. To verify the superiority of the proposed method, extensive experiments on t...