面向多尺度目标检测的改进Faster R-CNN算法

Improved Faster R-CNN for Multi-Scale Object Detection

由于多尺度目标检测中图像目标尺度差异性大,基于单层次特征提取的目标检测算法或者导致小目标特征提取丢失、扭曲,或者导致大目标特征提取冗余度过高,检测效果不理想.为此,基于Faster R-CNN 思想,提出一种多尺度目标检测算法.首先采用多层次提取特征策略提取多尺度目标特征;然后统计目标真实框大小与纵横比,设置锚点规格;最后采用多通道方法生成多尺度目标候选框.基于PASCAL VOC 数据集的实验结果表明,该算法总体漏检率为9.7%,平均精度的均值为75.2%,检测性能较当前主流的多尺度目标检测算法有一定的提高.

For multi-scale object detection, the detection methods based on single-level feature extraction suffered from the low detection quality because of the loss or distortion of feature for small-scale objects, or the redundancy of feature for large-scale objects. We propose a multi-scale object detection method based on Faster R-CNN. The method extracts the multi-scale features with the policy of multi-level feature extraction, configures statistically the size and the aspect ratio of the anchor, and adopts a multi-channel region strategy to generate multi-scale proposals. Extensive experiments on the PASCAL VOC dataset show that the quality of our method, with 9.7% of the log-average miss rate and 75.2% of the mean average precision, performs better than the traditional detection methods.