基于余弦注意力机制的少样本历史建筑识别方法

Few-Shot Historical Building Recognition Method Based on Cosine Attention Mechanism

建筑物作为城市空间中的核心地标和人类活动的重要载体,其精准识别在城市规划、智慧旅游等领域具有重要意义,然而收集和标注足够量的数据是一项昂贵且耗时的任务。针对部分建筑物标注数据稀缺且视觉外观多样性导致特征表示不足的问题,提出一种基于余弦注意力机制的少样本建筑识别方法。该方法利用自适应原型表示方法充分捕捉目标对象的特征,并使用余弦注意力机制代替Transformer中的缩放点积注意力机制以优化模型性能。首先收集来自公开资源的样本数据,其次构建了一个包含多种青岛历史建筑的少样本分类数据集,再次使用该数据集验证所提出方法的有效性。实验结果表明,在1-shot和5-shot学习场景中,该方法准确率分别达到了58.08%、77.15%,验证了该方法对少样本建筑的识别能力和效果。

As the core landmark in urban space and a crucial medium for human activities,the accurate identification of buildings is of great significance in urban planning,smart tourism and other fields.However,collecting and labeling a sufficient amount of data is a costly and time-consuming endeavor.Aiming at the problem that some buildings'labeling data are scarce and the visual appearance diversity leads to in⁃sufficient feature representation,this paper proposes a method of building identification with few samples based on cosine attention mecha⁃nism.This method fully captures the features of the target object by using an adaptive prototype representation methods,and replace the scaled dot-product attention mechanism in Transformer with a cosine attention mechanism to optimize model performance.Firstly,the paper collects sample data from open resources,then constructs a small sample classification data set containing a variety of historical buildings in Qingdao,and then uses this data set to verify the effectiveness of the proposed method.Experimental results show that the proposed method achieves 58.08%and 77.15%accuracy in 1-shot and 5-shot learning scenarios,respectively,which shows the ability and effect of this method in building identification with few samples.