【摘要】 针对使用Vgg-16作为特征提取网络的目标检测算法SSD存在的模型不够轻量化以及对弱小目标的检测能力较弱等问题，提出替换SSD主干网络和在其分类层前添加特征融合模块的改进方法。在对目标检测数据集Pascal VOC07+12进行预处理之后，选择轻量化特征提取网络PeleeNet作为SSD的主干网络，平均精度提高1.7%的同时也将模型权重缩减为原来的1/5。此外，在模型分类层前将SSD生成的前三个特征图进行特征融合，丰富了小目标特征图的背景信息，在VOC2007测试集上的平均精度再次提高了3.1%。同时，模型在VOC小目标类别上的平均精度也要比对比方法高出1～5.5个百分点。最后，将模型在嵌入式开发板Jetson TX2上进行部署运行，实现了深度学习模型在嵌入式终端上的部署与应用。更多还原

【Abstract】 Aiming at the problems that SSD,which used Vgg-16 as a feature extraction network, had huge computational cost and weak ability to detect small objects, the method of replacing the SSD backbone network and adding a feature fusion module in front of its classification layer was used in this paper. After pre-processing the object detection data set Pascal VOC07+12,the pre-trained lightweight feature extraction network PeleeNet was selected as the backbone of SSD,which had the mean Average Precision(mAP) increased by 1.7% and reduced the model size to one-fifth of the original. In addition, to enrich the background information of small object feature maps, the first three feature maps generated by SSD were fused before the model classification layer. The results on the VOC2007 test dataset show that the mAP increases by 3.1% again and the average accuracy of the model is also 1～5.5 percentage points higher than that of the comparison method on small object classes. Finally, the model was deployed to the embedded terminal Jetson TX2. It realizes the deployment and application of the deep learning model on the embedded terminal.更多还原