The customized parts in the construction machinery scene have complex shapes and wide-size spans. There is still room for improvement in the current research’s adsorption accuracy and time cost. To address these issues, this paper proposed a three-stage adsorption pose search method based on DenseNet acceleration and the GWO (gray wolf optimization). First, the DenseNet-121 is chosen to classify the synthesized part image. Then, the GWO is used to search for the adsorption pose of the part classified as irregular type. Finally, we store the 1024 vector and the corresponding adsorption pose as key-value data in the database for new part image retrieval. The experiments show that the average classification accuracy is as high as 99.3%, while its time cost is 487 times lower than the center of gravity rotation search. Moreover, the proposed GWO algorithm has improved the adsorption score by 7.7% compared with the state of the art (SOTA), which fully demonstrates the speed and reliability of our method.