加权聚合深度卷积特征的图像检索方法

An Image Retrieval Method Based on Weighting Aggregation of Deep Convolutional Features

针对目前基于深度卷积特征的图像检索方法无法充分突出图像显著性区域特征和不能有效抑制背景噪声等问题,提出了一种加权聚合深度卷积特征的图像检索方法.根据逆文档频率,该方法对拥有较少特征和紧密特征的特征图赋予较大权重,生成差异性加权向量.由于不同图像表现的特征不同,该方法选择最能真实反映图像特征的一组特征图,计算出权重矩阵并对其进行滤波处理,最终生成选择滤波加权矩阵.公开数据集上的实验结果表明,本文提出的方法能够有效地增强图像特征的辨别能力,在图像检索精度上优于其它同类方法.

As the current image retrieval methods based on deep convolutional features cannot fully highlight salient regional features of images and cannot effectively suppress background noises,an efficient unsupervised image retrieval method based on the weighting aggregation of deep convolutional features is proposed.According to the inverse document frequency,this method assigns a larger channel weighting to the feature map with fewer features and compact features and then generates a differentiated weighting vector.As different images have different features,this method selects a set of feature maps that can most truly indicate image features,calculates the weighting matrix,and performs a filtering process to generate a selected filter weighting matrix.Experimental results on different public datasets show that the proposed method can effectively improve the discrimination power of image features.Furthermore,under the same experimental setting,the proposed method is superior to other similar aggregation approaches on image retrieval accuracy.