Compared with traditional features, the features extracted by convolutional neural networks have a stronger image description capability, and the  convolutional layer is more suitable for image retrieval than the full connected layer. However, convolution features which are highdimensional consume huge time and memory to do image matching. We propose a new method to improve and integrate convolution features to construct a singledimensional feature vector, and then use it for image matching. Firstly, the 3D feature extracted from the last convolution layer is reweighted to highlight the edge information and location information of the image. Then, we integrate several regional feature vectors which are processed by sliding windows, into a global feature vector. Finally, we get the initial ranking results of the search by using the cosine distance to measure the similarity between the query graph and the test image, and conduct reranking with the extended query method to get the final mean average precision (mAP ). We conduct experiments on the Paris6k and Oxford5k  databases and two extended databases Paris106k and Oxford105k which are extended by100k images. On Paris databases, the proposed method improves the mAP by about 3% compared with the crossdimensional weighting features (CroW) method on Paris. On Oxford databases, our method is about 1%  better than the CroW method. The results show that the global features extracted by our new method can better describe the image.