基于编码-解码技术的图像标题生成方法分析

Analysis of image title generation method based on encoding decoding technology

文章首先介绍了图像标题生成的常用方法,包括模板法、检索法、编码-解码法;其次,在互助双向LSTM模型的基础上,详细介绍了图像标题生成算法的实现步骤,即利用Inception-V4编码器将原始图像编码成图像特征,并使用互助双向LSTM解码器将这些特征解码成相应的句子,同时采用语音混沌保密通信技术确保信息安全;最后,进行了实验测试,并通过多模态注意力可视化分析验证了递进解码机制的作用。实验结果显示,在LSTM解码机制的支持下,能够生成优质、精确的图像标题。

Firstly,this paper introduces the common methods of image title generation,including template method,retrieval method and encoding decoding method.Secondly,based on the mutual aid two-way LSTM model,the implementation steps of the image title generation algorithm are introduced in detail,that is,the original image is encoded into image features by using the perception-v4 encoder,and these features are decoded into corresponding sentences by using the mutual aid two-way LSTM decoder.At the same time,the speech chaotic secure communication technology is used to ensure information security.Finally,experimental tests are carried out,and the role of progressive decoding mechanism is verified by multimodal attention visualization analysis.Experimental results show that with the support of LSTM decoding mechanism,it can generate high-quality and accurate image titles.