Object detection based on SSD-ResNet

Abstract

Nowadays, with the abundance and diversity of the data sets of the detected objects, detection and recognition technology has achieved excellent performance in learning effect. However, because the target objects are usually very small in many real-world applications while the background environment seldom varies, manually annotating these objects is extremely costly in time and manpower. These problems have challenged the learning effect of standard neural networks. In this paper, we propose a novel method to replace the original network structure and to extend the number of layers for detecting many kinds of dangerous goods among different background environments. Specifically, we employ SSD as the basic network structure and replace the inside VGG16 with a ResNet101 network. The experimental results show that the ResNet network is effective in detecting many kinds of dangerous objects in small data sets. The proposed model outperforms other neural networks in learning efficiency and accuracy.