一种用于图像检索的幂归一化深度卷积特征加权聚合方法

Power-normalized weighting aggregation of deep convolutional features for image retrieval

针对图像检索中基于部位的加权聚合(PWA)方法存在的视觉突发问题,提出一种幂归一化的深度卷积特征加权聚合方法。首先简化了原PWA方法中用于确定空间权重的归一化和幂变换操作,直接将所选择的有区分性的通道特征图作为空间权重矩阵,然后引入新的幂变换函数并选取合适的参数对加权聚合后的通道响应进行归一化处理,最后通过PCA降维和白化处理形成图像的全局特征表示形式。在4个标准数据库上的图像检索实验结果表明,该方法能有效调节PWA聚合特征响应的突发度并提高图像检索的准确率。

To solve the problem of visual burstiness in the part-based weighting aggregation(PWA)method for image retrieval,a power-normalized weighting aggregation method for deep convolutional features is proposed.Firstly,the original normalization and power transformation for determining spatial weights in PWA method is simplified and the discriminative feature maps of the selected channels are directly used as the spatial weighting matrix.Then,a new power transformation function is introduced and the appropriate parameters are chosen to normalize the aggregated feature responses.Finally,the global feature representation of an image is obtained by PCA compression and whitening operations.The results of image retrieval experiments on four standard databases indicate that the proposed method can effectively regulate the burstiness of feature responses after PWA aggregation and greatly improve the image retrieval performance.