New default box strategy of ssd for small target detection

Abstract

SSD, which combines the advantages of Faster-RCNN and YOLO, has excellent performance in both detection speed and precision by merging the default boxes of six different layers. As the original default box strategy cannot accurately capture the small target information, the detection precision of SSD for small target images is not as good as normal size targets. In this paper, a new default box strategy, which can give the appropriate size and number of default boxes, is proposed to improve the performance of SSD for small target detection. The new default box strategy is made up of new scales and new aspect ratios. The new scales, which provide the basic scales for the six layers, are defined by the size ratio of the kernel to the convolutional layer. In addition, the new scale range is reduced from [20, 90] to [20, 60]. The new aspect ratios, which determine the size and the number of default boxes of the six layers, are defined as [[1.1], [1.1], [1.1], [1.1], [0.8, 1.2], [1.1]]. Experiment results on the small ground target dataset show that the detection precision of SSD with the new strategy is 99.5 mAP, which is 4.6 mAP higher than that of the original SSD. More importantly, the training time of SSD with the new strategy is 963 s or 326 s less than that of the original SSD.