基于改进PVANet的实时小目标检测方法

现有目标检测算法主要以图像中的大目标作为研究对象，针对小目标的研究比较少且存在检测精确度低、无法满足实时性要求的问题，基于此，提出一种基于深度学习目标检测框架PVANet的实时小目标检测方法。首先，构建一个专用于小目标检测的基准数据集，它包含的目标在一幅图像中的占比非常小且存在截断、遮挡等干扰，可以更好地评估小目标检测方法的优劣；其次，结合区域建议网络（RPN）提出一种生成高质量小目标候选框的方法以提高算法的检测精确度和速度；选用step和inv两种新的学习率策略以改善模型性能，进一步提升检测精确度。在构建的小目标数据集上，相比原PVANet算法平均检测精确度提高了10.67%，速度提升了约30%。实验结果表明，该方法是一个有效的小目标检测算法，达到了实时检测的效果。

Real-time small object detection method based on improved PVANNet

Existing object detection algorithms are mainly aimed at detecting big objects in an image. Research on small object detection is still too scarce and there are problems with low detection accuracy and failure to meet the real-time requirement. This paper proposed a real-time small object detection method based on deep learning framework PVANet. Firstly, it built a benchmark dataset especially for small object detection problem. The dataset consisted of small objects covering a very small part of an image and also contained some interferences such as truncation and overlap. Secondly, combining with the region proposal network(RPN), it designed a strategy to generate high-quality candidate proposals for small objects to raise the detection accuracy and speed. Finally, it adopted two new learning rate policies step and inv to further enhance the detection accuracy. The proposed method achieved the mAP(mean average precision) by 10.67% and speed by 30% improvement over the original PVANet algorithm. Experimental results show that this method is effective on small object detection and can run in real-time.