Scholarlink

Generative AI enables the creation of synthetic data to train AI models, offering new possibilities for expanding training datasets and overcoming data limitations. However, generating images for Visual Question Answering (VQA) is challenging, as the augmented images must preserve the semantic consistency of the \(&lt;image_ question_="" answer=""&gt;\) triplet. This study investigates the use of image generation models to augment data and enhance VQA performance. An automated augmentation and evaluation process is designed, employing prompt engineering to generate effective prompts for the image generation process. The generated images are evaluated using the proposed assessment method, and the VQA model is trained on both original and augmented datasets to assess effectiveness. Results demonstrate that training on the augmented dataset improves model performance, showcasing the value of learning from augmented image features. These findings also emphasize the importance of innovative data augmentation techniques in enhancing AI models, particularly for tasks that require a deep understanding of multimodal information. The code is available at https://github.com/phuthinhnhpt123/open-ended-VQA .</image_>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing Visual Question Answering with Semantic-Preserving Image Generation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Phu-Thinh Nguyen-Huynh,
  • Huy Tien Nguyen,
  • Tung Le

摘要

Generative AI enables the creation of synthetic data to train AI models, offering new possibilities for expanding training datasets and overcoming data limitations. However, generating images for Visual Question Answering (VQA) is challenging, as the augmented images must preserve the semantic consistency of the \(<image_ question_="" answer="">\) triplet. This study investigates the use of image generation models to augment data and enhance VQA performance. An automated augmentation and evaluation process is designed, employing prompt engineering to generate effective prompts for the image generation process. The generated images are evaluated using the proposed assessment method, and the VQA model is trained on both original and augmented datasets to assess effectiveness. Results demonstrate that training on the augmented dataset improves model performance, showcasing the value of learning from augmented image features. These findings also emphasize the importance of innovative data augmentation techniques in enhancing AI models, particularly for tasks that require a deep understanding of multimodal information. The code is available at https://github.com/phuthinhnhpt123/open-ended-VQA .