Abstract:

As an important research direction of computer vision, object detection plays an increasingly important role in smart cities and unmanned driving. In traditional object detection algorithms, it judges positive and negative samples according to the threshold value of the Intersection over Union（IOU）. Some lower IOU will introduce noise and reduce the accuracy of the detector, and higher IOU will retain some high-quality samples, resulting in over-fitting. The difference between the IOU threshold of the recommended area and the detector will cause quality mismatch. In response to these problems, this paper proposes a parallel cascaded detection network based on cascaded network, which is composed of various of detectors connected in series and in parallel. Every of each detector sets an incremental IOU threshold, thus at each stage a higher quality distribution is obtained to train the next detector and gradually resample to reduce overfitting. The experimental results show that the proposed parallel cascaded detection network is better than the traditional object detection algorithm in detection precision. The average Accuracy Precision（AP） of the object detection dataset Microsoft COCO is increased by about 1.5 percentage points using the proposed parallel detection network.

Key words: convolutional neural network, deep learning, cascaded network, high-precision object detection

摘要：

目标检测作为计算机视觉的重要研究方向，在智慧城市、无人驾驶等领域的作用越来越重要。传统目标检测算法中，根据交并比（Intersection over Union，IOU）的大小判断正负样本，但较低的IOU会引入噪声，降低检测器的精度；较高的IOU会保留少数高质量样本，造成过拟合；并且推荐区域和检测器的IOU阈值相差过大会引起质量不匹配问题。针对上述问题，提出了一种基于级联网络的平行级联检测网络，它由一系列检测器串并联而成，每个检测器设置递增的IOU阈值，从而在每个阶段都会得到一个更高质量的样本分布来训练下一级检测器，并逐步重采样减少过拟合。实验结果表明提出的平行级联检测网络的检测精度优于传统目标检测算法，在目标检测数据集Microsoft COCO上平均准确度（AP）提升了1.5个百分点左右。

关键词: 卷积神经网络, 深度学习, 级联网络, 高精度目标检测