DMVMLC-VT: Deep incomplete multi-view multi-label image classification with view translation and pseudo-label enhancement

In recent years, the problem of multi-view multi-label classification has attracted widespread research interest. However, in practical applications, multi-view multi-label data often grapple with issues such as data acquisition errors and human annotation errors, compromising the completeness of both views and labels. To tackle the challenge of incomplete multi-view and incomplete multi-label classification, we propose a deep learning approach, DMVMLC-VT, based on view translation and pseudo-label enhancement. Distinct from previous dual prediction methods, our view translation-based completion module can accommodate representations with missing data as input, enhancing the utilization of multi-view data with missing components. We employ autoencoders to learn specific view representations from raw data, subsequently utilizing a completion module to refine these representations. A contrastive fusion strategy is adopted to bolster the discriminative power of the common representations. Lastly, we introduce pseudo-labels and a label balancing strategy to optimize the utilization of data with missing labels, thereby achieving superior classification performance. Experimental results on five datasets show that our method outperforms five other competing methods. The code could be accessed from https://github.com/Lyc1022/DMVMLC-VT.