【摘要】 针对基于区域的目标检测算法中定位精度不高的问题,本文提出了一种分裂机制的改进Faster RCNN算法。该算法首先选择特征提取能力强的卷积神经网络(CNN)作为骨干网络进行特征的提取;然后通过12种不同Anchors产生候选目标区,以进一步提升检测的精确度;最后将得到的特征分别传送到两个子网络,分别实现分类与定位。分类网络以全连接结构为基础,定位网络则主要由卷积神经网络构成。本文在Pascal VOC2007和Pascal VOC2012以及吸尘袋图像集上对算法的有效性进行了验证。结果表明,提出的算法在对目标进行有效检测的同时,定位效果比Faster RCNN更加精确,实现了边界框的精准回归。更多还原

【Abstract】 For the region-based target detection algorithm,there is a general problem that the boundary position is far from the real value.This paper proposes an improved Faster RCNN algorithm using the splitting mechanism.Firstly,the algorithm selects the convolutional neural network(CNN) with strong feature extraction ability as the backbone network to extract features,and then generates candidate target regions through 12 different anchors to further improve the accuracy of detection.Finally,the obtained features are transmitted to two different sub-networks:the classification network is based on the fully-connected structure,and the targets are classified;the positioning network is based on the convolutional neural network structure to achieve the target positioning.The experiments verify the effectiveness of the algorithm on the Pascal VOC2007 dataset,Pascal VOC2012 dataset and vacuum bag dataset.The results show that the proposed algorithm is more accurate than the Faster RCNN in the effective detection of the target,and achieves the accurate regression of the bounding box.更多还原