fig 1.docx

Transformer-based neural network models have demonstrated remarkable success in handling both text generation tasks and multimodal tasks, such as generating image captions. Over time, advancements in these models have evolved from traditional encoder-decoder architectures to transformer-based systems utilizing vision-language pre-training (VLP). Studies reveal that pre-training significantly enhances the performance of these models, making them highly effective for complex descriptive text generation tasks. However, challenges remain regarding the inclusion of crucial factual information required for accurate comprehension of visual content by individuals with visual impairments. Current models tend to overlook details such as the authorship, dimensions, and stylistic elements of artwork, highlighting the necessity for further refinement of teaching methodologies and model tuning. This paper explores the development and application of transformer models in audio descriptive text generation while discussing key strategies to enhance their performance.