基于改进的特征提取网络的目标检测算法

Object Detection Algorithm Based on Improved Feature Extraction Network

针对目标检测准确率低,物体位置不精准的缺点,设计了一种基于改进的特征提取网络的目标检测算法。首先将训练集进行数据增强;其次设计了一种双通道网络,用于目标检测算法Faster R-CNN的特征提取;最后在算法的预测部分,对非极大值抑制(NMS)机制进行了改进,并采用加权平均方法获取存在多个相近的预测框的位置。在VOC 2007和VOC 2012数据库上进行实验,表明本文算法比经典的目标检测算法效果要好,准确率达到79.1%,提升了3%~4%,验证了本文算法的有效性。

In this study,an object detection algorithm is designed based on an improved feature extraction network to solve the shortcomings of low object detection accuracy and inaccurate object position detection.Initially,the training set is enhanced;subsequently,a two-path network is designed for usage in feature extraction of the Faster R-CNN algorithm;finally,the non-maximum suppression mechanism is improved in the prediction part of the algorithm,and the weighted averaging method is adopted for obtaining the positions of multiple similar prediction boxes.The experiments conducted using the VOC 2007 and VOC 2012 databases denote that the proposed algorithm outperforms the classical object detection algorithm,with an accuracy rate of 79.1% and an improvement of 3%-4%.Thus,the effectiveness of the algorithm is verified.