According to the problem of unsatisfied accuracy of image retrieval based on image feature and ag-gregation coding, a new image description algorithm RCroW based on multi-region cross-weighted aggrega-tion deep convolution feature is proposed. Firstly, the target region is vaguely labeled because the target region has the characteristics of high activation response. The spatial weighted matrix is generated by combining the feature map and the mask map. Then, multi regional strategy is introduced to transform the spatial weighted matrix into multi-regional cross-weighted matrix. Finally, the image feature vector is generated by multi-region cross-weighted matrix aggregating deep convolution feature. Experiments were carried out on three datasets of Oxford5k, Paris6k and Holidays. And the experimental results show that the average accuracy of image re-trieval based on image feature description algorithm RCroW is better than the other seven commonly used al-gorithms such as CroW, R-MAC, SPoC and so on.%针对依赖图像特征和聚合编码的"以图搜图"方法检索准确率较低的问题, 提出一种基于多区域的交叉加权聚合深度卷积特征描述算法——RCroW. 首先利用目标区域具有较高激活响应的特性标记出目标轮廓位置, 将卷积特征图和目标轮廓掩码图结合生成空间权重矩阵; 然后引入多区域策略, 将空间权重矩阵转变成多区域交叉权重矩阵; 最后利用多区域交叉权重矩阵加权聚合深度卷积特征生成图像特征向量. 在 Oxford5k, Paris6k 和 Holidays 这3个数据集上进行的实验的结果表明, RCroW算法的图像检索平均准确率优于CroW, R-MAC和SPoC等7种算法.
展开▼
