【摘要】 针对现有图像描述生成算法在解码阶段由于语言模型结构简单,解码表达能力较弱,容易造成语义缺失的问题,引入有序长短时记忆网络(ON-LSTM),改进现有模型解码器,构建双层LSTM架构,显式的提取描述文本层级结构,解码出更丰富的语义特征。在MSCOCO数据集上进行训练和测试,实验结果表明,改进的算法能够生成更加符合自然语言习惯的描述语句。更多还原

【Abstract】 Aiming at the existing image description generation algorithm,in the decoding stage,the language model is simple in structure and weak in decoding expression ability,which can easily cause the problem of lack of semantics.The ordered neurons Long Short-Term Memory network(ON-LSTM)is introduced to construct a two-layer LSTM architecture to improve the decoder of the existing model,so that it can explicitly extract the text hierarchical structure of the description to decode richer semantic features.Training and testing on the MSCOCO data set,the experimental results show that the improved algorithm can generate description sentences that are more in line with natural language habits.ordered neurons Long Short-Term Memory network更多还原