基于改进注意力机制的图像描述算法

Image description algorithm based on improved attention mechanism

图像描述的任务是根据输入图像自动生成描述该图像的语句,属于计算机视觉与自然语言处理的交叉领域。针对传统注意力机制提取特征能力不足、模型复杂且训练困难等问题,本文提出了一种改进注意力机制的图像描述模型。在传统注意力机制的基础上引入高效通道注意模块,在提升特征提取效果的同时降低模型复杂度,在保证性能的同时提高模型效率,更好的提取图像重要部分特征,生成更为准确的自然语言描述。模型在MSCOCO数据集上进行了验证,实验结果表明,相较于传统的注意力机制,模型在生成描述语句准确性方面有较大提升,在BLEU-1、BLEU-3、BLEU-4上分别有0.2%、0.3%、0.6%的提高。

Image description task uses machine to automatically generate the description of the image according to the input image,which belongs to the intersection of computer vision and natural language processing.In order to solve the problems of traditional attention mechanism,such as insufficient feature extraction ability,complex model and difficult training,this paper proposes an improved image description model of attention mechanism.Based on the traditional attention mechanism,the efficient channel attention module is introduced to improve the feature extraction effect while reducing the model complexity,improve the model efficiency and ensuring the performance results,better extraction of the features of important parts of the image and more accurate generation of natural language description.The model is validated on MSCOCO data set.Experimental results show that compared with the traditional attention mechanism,the model has a significant improvement in the accuracy of description generation,with an improvement of 0.2%in BleU-1,0.3%in BleU-3 and 0.6%in BLEU-4,respectively.