Automatic Descriptive Text Generation Method for Sports Scenes Using Dilated Convolutional Neural Networks and BERT Pretraining Model

The automatic conversion of sports videos or image content to natural language description remains significant, which provides powerful tools for sports analysis, video annotation and disability assistance. In the context, this paper proposes an automatic generation method for motion scene description text based on dilated convolutional networks (DCNN) and Bidirectional Encoder Representation from Transformers (BERT) model. The study first uses Faster R-CNN to detect the regions of motion scene images and standardizes the feature maps through ROI pooling. Then, using ASPP technology, DCNN efficiently extracts key visual features from images at multiple scales and integrates BERT with geometric attention for encoding and Transformer decoding. Finally, through a two-stage training strategy, the extracted visual features are transformed into coherent and accurate natural language descriptions, thereby achieving effective alignment and fusion between visual features and textual descriptions. The experimental results show that this method has achieved significant performance improvement on multiple motion scene datasets, and the generated descriptive text is not only rich in content and accurate in details but also has high fluency and readability.