基于RPR-Transformer图像描述生成模型  

Image Description Generation Model Based on RPR Transformer

作　　者：赵芸 ZHAO Yun(Shanghai Baoxin Software Co.,Ltd.,Shanghai 200000)

机构地区：[1]上海宝信软件股份有限公司,上海200000

出　　处：《长江信息通信》2024年第12期84-86,共3页Changjiang Information & Communications

摘　　要：图像描述生成结合了计算机视觉和自然语言处理,旨在为图像提供准确描述。注意力机制忽略了图像的二维空间特性。文章提出基于物体间相对位置关系的自注意力模型(RPR-Transformer)。通过目标检测技术提取物体特征并计算对应物体的中心位置以及面积;使用关系特征提取模型提取图像中物体之间的关联特征;对融合后的特征使用门控单元过滤,去除干扰信息。实验结果表明本模型具有较强的鲁棒性。Image description generation combines computer vision and natural language processing,aiming to provide accurate descriptions for images.The attention mechanism ignores the two-dimensional spatial characteristics of images.This article proposes a self attention model(RPR Transformer)based on the relative position relationship between objects.Extract object features through object detection technology and calculate the center position and area of the corresponding object;Using a relational feature extraction model to extract the correlation features between objects in an image;Filter the fused features using gating units to remove interference information.The experimental results indicate that this model has strong robustness.

关 键 词：图像描述生成 关系特征提取 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]