<p>Current image captioning methods based on grid features and segmentation features have achieved some success. However, existing methods combine the two visual features by directly fusing them, which inevitably introduces semantic noise. In order to achieve more effective feature fusion, we propose a <i>Visual-Text Information Enhancement Network</i> (<b>VTIENet</b>) to enhance the information representation of visual and text in the encoding and decoding stages, respectively. Specifically, in the visual feature encoding stage, we propose the Visual Enhancement Encoder (VEE), which uses the segmentation features as the visual complementary information of the grid features, and realizes the deep fusion of the two visual features through Adaptive Visual Fusion Attention (AVFA) to obtain a richer visual feature representation. In the language decoding stage, we propose a Text Enhancement Decoder(TED), which uses a pre-trained context encoder to encode the generated caption information, and uses Context Guided Self-Attention (CGSA) to aggregate the encoded sentence-level information to guide the decoder to generate more accurate words at each time step. To validate our model, we conduct a large number of experiments and visualizations on MS-COCO dataset. The experimental results show that VTIENet is competitive with existing models in performance.</p>