基于改进Faster RCNN的目标检测算法

An object detection algorithm based on improved Faster RCNN

针对Faster RCNN算法在对小目标、发生形变的目标以及重叠度大的目标检测存在的问题,做了相应的改进。将具有更深层次且表达能力更强的ResNeXt网络代替了Faster RCNN中用于提取图像特征的VGG-16网络。将ResNeXt常规的“后激活”基本单元结构更改为“预激活”的形式,反向传播过程中,信息可以在整个网络中“直接”传播,使得网络的性能达到最优。引入可变形卷积,解决产生空间形变的图像识别任务。对比实验结果表明,本文设计的网络模型在对发生旋转以及局部遮挡等状况的目标表现出良好的检测效果和较高的准确性。

We propose some improvements to solve the existing problems of applying the Faster RCNN algo-rithm to small targets,deformed targets,and large overlap targets.The VGG-16 network in Faster RCNN is replaced by the ResNeXt network with a deeper level and a stronger expression ability in extracting image fea-tures.First,the basic unit structure of ResNeXt is changed from"post-activation"to"pre-activation",so that the information can be propagated"directly"in the whole network during back propagation and the network performance can be optimized.Secondly,a deformable convolution is introduced to recognize images with spatial deformation.Lastly,to solve the problem of missing detection caused by target overlap,the softNMS algorithm is introduced to improve the screening method of the candidate frames.The experimental results show that the proposed model has good detection ability with high accuracy for rotated and partially blocked targets.