联合加权聚合深度卷积特征的图像检索方法

An image retrieval method based on deep convolutional features of joint weighting aggregation is proposed to solve the problem that most existing image retrieval approaches can't fully extract image features and their performance requires to be improved. Firstly, the method extracts the outputs of the last convolutional layer as deep convolutional features of an image by passing the image through a pre-trained deep convolutional neural network. Then, the spatial weight matrix is calculated to highlight significance regions of the image and to suppress the background noise of the image. The maximum-principle of channel variance is then used to select the corresponding feature map and to calculate the spatial weight matrix. The original deep convolutional features are weighted and aggregated into a feature vector. Moreover, the channel weight vector is calculated by distinguishing feature maps of different channels, and then the global feature representation of this image is obtained by multiplying the aggregated feature vector and the channel weight. Experimental results on different public available datasets for image retrieval show that the proposed approach effectively enhances the discriminative ability of image features, outperforms the state-of-the-art approaches based on pre-trained networks and can be effectively applied to related fields of image retrieval.