In order to solve three problems of traditional recommendation models, i.e., data sparsity, low robustness and the lack of deep-level semantics among heterogeneous features, a novel correlation visual adversarial Bayesian personalized ranking (CVABPR) recommendation model was proposed. First, based on the movie titles in the original MovieLens datasets, the corresponding movie posters were downloaded from Internet movie database (IMDB) to construct two multimodal datasets named MovieLens-100k-WMI and MovieLens-1M-WMI, respectively. Second, a group of heterogeneous but complementary image features were extracted using the SENet model to describe movie posters accurately. Then, the cluster canonical correlation analysis model was improved to mine the implicit cluster canonical correlation between the heterogeneous features. Afterwards, the correlation was used to optimize the visual Bayesian personalized ranking (VBPR) model to better depict the movies to be recommended. Finally, a perturbation factor was absorbed into the recommendation model to enhance the robustness of the CVABPR model through adversarial learning, making the recommendation model more stable and generating high-quality recommendation results. To verify the proposed CVABPR model, a set of experiments were carried out on two multimodal datasets. Evident performance improvements of the CVABPR model were observed on the two datasets. Specifically, a 3.802% performance improvement of the mean average precision (MAP) metric was obtained on the MovieLens-100k-WMI dataset, and a 4.609% performance improvement of the MAP metric was observed on the MovieLens-1M-WMI dataset. The mainstream baseline was defeated by the CVABPR model. Based on ablative analysis experiments, a more important role of the adversarial learning strategy was found compared with the cluster canonical correlation. Additionally, larger performance improvements were observed on the MovieLens-1M-WMI dataset with higher data sparsity. The key challenges of data sparsity and the lack of deep semantic among heterogeneous features were solved to a certain degree. Meanwhile, the CVABPR model has strong robustness