SSD(Single Shot MultiBox Detector)目标检测算法的研究与改进

目标检测是图像处理的重要领域,SSD算法是基于深度学习下对物体进行检测。首先对SSD检测算法进行了深入研究,然后使用MxNet框架对SSD算法进行了实验,并对训练过程m AP值以及分类准确率进行了分析。由于SSD算法对小物体的检测准确率并没有那么理想,通过在预测层加入反卷积和对预测框长宽比的调整,使得改进后物体检测的平均准确率得到了可观的提高,并且提高了对小物体的检测准确率。

Firstly,the SSD detection algorithm is deeply studied in this paper.Then the SSD algorithm is tested by MxNet framework,and the training process mAP value and classification accuracy are analyzed in this paper.Since the accuracy of SSD algorithm for small objects is not so good,by adding deconvolution in the prediction layer and adjusting the aspect ratio of the prediction frame,the average accuracy of the improved object detection is greatly improved and improve the accuracy of detection of small objects.