Controllable Image Caption Generation Based on Encoder-decoder for Power Construction Scene

Image caption generation of electric power construction scene adopts deep learning based on encoding and decoding technology to understand the image information and convert it into text description output, so as to warn the potential security risks and enrich the output forms of traditional image analysis technology. The traditional image caption generation method lacks controllability and has insufficient detail descriptions, and there are few the researches on image description of electric power construction scene. Therefore, an optimization method of controllable image caption generation based on encoding and decoding is proposed. A new feature extraction model, the FVC R-CNN, is introduced as an encoder to extract the salient features and common visual features of the images. An improved MT-LSTM network for feature decoding is obtained by improving the activation function. Finally, the output is optimized by a multi-branch decision strategy. The power scene description dataset is trained and tested on the Ubuntu16.04 and PyTorch deep learning framework. Experimental results show that the accuracy of image caption generation is significantly improved, and the controllability of scene description is enhanced, which effectively improves the intelligent level of safety management in the power construction scene. © 2022 Power System Technology Press. All rights reserved.