Optimal transport enhanced text-assisted discriminative clustering for intelligent network images analytics

In digital network, the uncontrolled flow of massive web image data poses a severe challenge to efficient organization and retrieval. Traditional deep clustering methods are often costly to train and may lack sufficient discriminative power to solve these problem. This paper proposes a model to support intelligent network image analysis: Optimal Transport enhanced discriminative text-Assisted Clustering (OTAC). The OTAC model aims to enhance the discriminative power of clustering models for intelligent network image analysis. OTAC first uses a pretrained vision-language model to extract image features to reduce training cost. OTAC then utilizes an improved optimal transport mechanism to select more discriminative words for network image data, and constructs a more distinguishing text space. Subsequently, it combines optimal transport with learnable prompts to more effectively focus on fine-grained features of images in digital networks. This design further significantly improves the model’s discriminative ability and clustering performance. Extensive experiments on 7 datasets demonstrate that OTAC significantly outperforms current state-of-the-art models, validating its effectiveness for intelligent network image clustering and robust downstream automated analytics.