Abstract: Existing object detection models often use multi-scale feature fusion of feature pyramid to improve small object detection performance. However, in the shallow feature layer of the feature pyramid, the detection of small objects will be weakened due to the existence of large objects, and the semantic information of the upper feature layer will be lost due to lateral connection. To solve the above problems, an I-FPN feature pyramid is proposed. At the shallow feature layer, it erases the big object information and makes the model focus more on the small object. In the upper feature layer, the residual feature enhancement module is used to reduce the information loss. In addition, the model uses data augmentation techniques to improve the robustness. I-FPN feature pyramid was tested on VEDAI small target data set and PASCAL VOC universal target data set using Resnet master network. The experiment shows that, under the condition that the detection speed is not affected, the mAP index of VEDAI test set is increased by 2.4%, and that of VOC test set is increased by 0.5%.

Key words: object detection, deep learning, feature pyramid, small object