摘要
对基于生成对抗网络的图像描述生成框架进行改进,提出一种用于图像描述的高效编码方法。在编码器部分,引入轻量级编码网络EfficientNet-B0,提取图像特征的同时降低模型复杂度。在解码器部分,使用参数量较少的门控循环单元(Gate Recurrent Unit,GRU)生成对应的图像描述。同时,引入Ranger优化器,加快模型优化和收敛。实验结果表明,该编码方法在保证生成描述多样性和准确性的同时,能更加高效地编码图像信息,降低训练时长。
Image captioning generation framework based on generative adversarial network is improved,an efficient coding method for image captioning is proposed.In the encoder part,a lightweight encoding network,EfficientNet-B0,is adopted to extract image features while reducing the model complexity.In the decoder part,gate recurrent unit(GRU)with fewer parameters is adopted to generate corresponding image captioning.An optimizer Ranger is introduced to speed up model optimization and convergence.The experiment results show that the method can encode image information more efficiently and reduce the training time while ensuring the diversity and accuracy of generated captioning.
作者
林椹尠
冯菲蓉
LIN Zhenxian;FENG Feirong(School of Science,Xi'an University of Post and Telecommunications,Xi'an 710121,China;School of Communications and Information Engineering,Xi'an University of Post and Telecommunications,Xi'an 710121,China)
出处
《西安邮电大学学报》
2022年第3期77-83,共7页
Journal of Xi’an University of Posts and Telecommunications