A Balanced Feature Fusion SSD for Object Detection

摘要

Single shot multibox detector (SSD) takes several feature layers for object detection, but each layer is used independently. This structure may ignore some context information and is not conducive to improving the detection accuracy of small objects. Moreover, the imbalances of samples and multi-tasks during SSD training process can lead to inefficient training and model degradation. In order to improve the performance of SSD, this paper proposes a balanced feature fusion SSD (BFSSD) algorithm. Firstly, a feature fusion module is proposed to fuse and refine different layers of the feature pyramid. Then, a more balanced L1 loss function is proposed to further solve these imbalances. Finally, our model is trained with Pascal VOC2007 and VOC2012 trainval datasets and tested on Pascal VOC2007 test datasets. Simulation results show that, for the input size of 300 × 300, BFSSD exceeds the best results provided by the conventional SSD and other advanced object detection algorithms.