Based on the analysis of the single shot multibox detector (SSD) target detection algorithm, we propose a deep multi-scale feature fusion target detection (DMSFFD) algorithm based on deep learning. The SSD feature layer and its adjacent layer are first fused and the 3 pixel×3 pixel convolution layer is added into the feature map after fusion to reduce the aliasing effect of upsample. Then the deeper feature fusion is conducted and the upsample operation is performed respectively for three small convolution layers. Subsequently the concate operation is performed for four feature layers to generate feature maps with richer semantic information, and thus the multi-scale small target detection is realized. In order to save computing resources and improve the real-time performance of the algorithm, VGG16 is selected as the basic network here. Although the fused algorithm is more complex than SSD, its real-time performance is basically guaranteed. Moreover, the DMSFFD algorithm can successfully detect the small targets missed by most SSD networks, and its detection accuracy is also greatly improved compared with that of SSD.