目标检测语义特征增强网络

Object Detection Semantic Feature Enhancement Network

针对目标检测算法中图像边缘细节等高频特征提取不充分而导致检测效果较差的问题,提出一种目标检测语义特征增强网络(SFENet)。利用八度卷积构建高频语义特征增强模块,通过调节卷积特征中高低频分量来增强图像高频特征提取,从而捕捉更多的图像边缘细节信息,进一步加强网络对全局语义特征的理解,有效提升目标检测精度。实验结果表明:SFENet在Pascal VOC07+12数据集和自制锥桶数据集上的检测结果与YOLOv4相比,平均精度(mAP)分别提高0.34%和0.37%,适应于自动驾驶和机器视觉领域,可有效应用于中国大学生无人驾驶方程式大赛的环境感知任务(FSAC)。

A semantic feature enhancement network(SFENet)for target detection is presented to address the issue that high-frequency features,including picture edge details,are not effectively extracted in the target detection algorithm,resulting in poor detection results.To capture more information about the image edge details and further strengthen the network's understanding of the global semantic features,this paper uses octave convolution to construct a high-frequency semantic feature enhancement module.By adjusting the high and low-frequency components in the convolutional features,this module improves the image's high-frequency feature extraction and effectively increases the target detection accuracy.The experimental results show that the detection results of SFENet on the Pascal VOC07+12 dataset and the self-built conical bucket dataset improve the mean accuracy(mAP)by 0.34%and 0.37%,respectively,compared with YOLOv4.This network can be adapted to the fields of autonomous driving and machine vision and effectively applied to the environment perception task of the Formula Student Autonomous China(FSAC).