In order to enrich the expression of single-view features and realize complementary learning between multiple views, a multi-view latent space fusion representation method based on enhancing feature deep learning was proposed. The model has three submodules: single-view enhanced learning, multi-view complementary fusion, and self-representation based on clustering task. First of all, the dynamic routing mechanism of capsule network is introduced, and the interval loss penalty item is added to the objective function to obtain the single-view feature of differential features enhancement. Next, the important features of different views are fused, the public latent space of multiple views is learned, the complementary representation between view features is realized, and the classification task is met. Then, the subspace clustering algorithm is used to learn the self-representation matrix of latent space, and the low-rank representation constraints of the latent space reconstruction error matrix and the noise data matrix are added to the objective function to obtain the fusion representation that meets the clustering task. Finally, the classification and clustering experiments were conducted on four data sets. Compared with multiple benchmark algorithms, this algorithm has been steadily improved in performace, and the learned fusion characterization can better meet the needs of downstream classification and clustering tasks.