Deconfounded fashion image captioning with transformer and multimodal retrieval

Abstract

Background

The annotation of fashion images is a significantly important task in the fashion industry as well as social media and e-commerce. However, owing to the complexity and diversity of fashion images, this task entails multiple challenges, including the lack of fine-grained captions and confounders caused by dataset bias. Specifically, confounders often cause models to learn spurious correlations, thereby reducing their generalization capabilities.

Method

In this work, we propose the Deconfounded Fashion Image Captioning (DFIC) framework, which first uses multimodal retrieval to enrich the predicted captions of clothing, and then constructs a detailed causal graph using causal inference in the decoder to perform deconfounding. Multimodal retrieval is used to obtain semantic words related to image features, which are input into the decoder as prompt words to enrich sentence descriptions. In the decoder, causal inference is applied to disentangle visual and semantic features while concurrently eliminating visual and language confounding.

Results

Overall, our method can not only effectively enrich the captions of target images, but also greatly reduce confounders caused by the dataset. To verify the effectiveness of the proposed framework, the model was experimentally verified using the FACAD dataset.