The use of deep neural networks to extract deep high-level features has provided excellent image retrieval performance. However, few studies have analyzed low-level features, which have their own advantages in image representation. Combining the advantages of low-level features and deep features remains a challenging problem. To solve it, we propose a deep feature aggregation method that includes texture features. Its highlights are: (1) A novel covariance weight is proposed to select some key feature maps. This enhances the discriminativeness of feature maps. (2) A novel response-region weighting scheme is proposed to balance various response-regions weights. It improves the positive role of some valuable feature maps in retrieval. (3) A novel method is proposed to extract texture features together with deep features. It can combine deep and texture features into a compact representation, and improve the performance of image retrieval. Experiments were conducted with five popular benchmark datasets to show that our method can significantly improve image retrieval performance.