基于反馈机制与空洞卷积的道路小目标检测网络

随着卷积神经网络与特征金字塔的发展，目标检测在大、中目标上取得了突破，但对于小目标存在漏检、检测精度低等问题。在YOLOv4算法的基础上进行改进，提出YOLOv4-RF算法，进一步提高模型对小目标的检测性能。使用空洞卷积替换YOLOv4中Neck部分的池化金字塔，在网络更深处减少语义丢失的同时获得更大的感受野。在此基础上，对主干网络进行轻量化并增加特征金字塔到主干网络的反馈机制，对来自浅层与深层融合的特征再次处理，保留更多小目标的特征信息，提高网络分类和定位的有效性。鉴于小目标物体属于困难检测样本，引入Focal Loss损失函数，增大困难样本的损失权重，形成YOLOv4-RF算法。在KITTI数据集上的实验数据表明，YOLOv4-RF在各个类别上的检测精度均高于YOLOv4，并在模型缩小138 MB的基础上提高了1.4%的平均精度均值(MAP@0.5)。

Road Small Target Detection Network Based on Feedback Mechanism and Dilated Convolution

With the development of Convolutional Neural Network(CNN) and feature pyramids, target detection has made breakthroughs in large and medium targets, but there are missed detections and low detection accuracies for small targets.Aiming at the reasons for less information of small targets in the picture and the difference in the size of small targets from that of large targets, this study proposes the YOLOv4-RF algorithm based on the YOLOv4 algorithm and further enhances the detection performance of the model for small targets.This study uses dilated convolution to replace the pooled pyramid of the neck in YOLOv4 to reduce semantic loss and obtain a larger receptive field in the deeper part of the network.Moreover, the backbone network is lightweight and a feedback mechanism from the feature pyramid to the backbone network is added.The features from shallow and deep fusion are processed again, which retains more feature information of small targets and improves the effectiveness of the network classification and positioning.Finally, because the small target object belongs to the difficult detection sample, the focal loss function is introduced to increase the weight loss of the difficult sample and form the YOLOv4-RF algorithm.The experimental data on the KITTI dataset show that the detection accuracy of YOLOv4-RF in each category is higher than that of YOLOv4, and the Mean Average Precision(MAP@0.5) is improved by 1.4% by reducing the model by 138 MB.