Research on an Efficient Single-Stage Multi-object Detection Algorithm

Abstract

To further improve the detection accuracy of SSD object detection algorithm, in this paper, a high efficient single shot multibit detector (HE-SSD) algorithm is proposed, which based on SSD for solving the low accuracy of classical single-stage object detection SSD algorithm. Firstly, an efficient and dense network is designed to improve the detection accuracy. Secondly, in order to improve the robustness of the algorithm and solve the problem of positive and negative sample imbalance in the detection process, the Focal Loss function is used to suppress the weight of the easily classified samples in the loss function. Finally, the accuracy of SSD algorithm for small object detection is improved by data augmentation. In the experiment, the network structure is deployed through the Pytorch deep learning framework, compared the effects of SGD and Adabound optimization methods on training loss to verify the superiority of convergence of the proposed algorithm. The experimental results show that HE-SSD algorithm is more accurate than SSD in PASCAL VOC dataset.