激光与光电子学进展, 2020, 57 (12): 121013, 网络出版: 2020-06-03
基于深度学习的目标检测与可行域分割研究 下载: 1005次
Research on Target Detection and Feasible Region Segmentation Based on Deep Learning
图像处理 共享网络 多尺度特征 多任务联合 深度学习 image processing shared network multi-scale feature multi-task coalition deep learning
摘要
为提高智能车快速检测物体对多种场景的适应能力,提出了多任务共享一个特征提取网络的联合方法。首先用ResNet-50作为编码器提取图像的特征;然后采用单发多框检测算法的多尺度特征预测和快速回归思想,对检测结果进行解码,采用DeepLab v3中的多孔空间金字塔池化结构,对经ResNet-50下采样后的图像特征进行多尺度映射、双线性上采样和批次归一化处理,完成分割解码;最后设定好参数训练联合方法。实验结果表明,该方法的平均精度均值为89.00%,分割平均交并比为83.0,每秒传输帧数为31 frame,满足智能车的应用需求。
Abstract
In order to improve the adaptability of intelligent vehicles to quickly detect objects in various scenes, a joint method of multi-task sharing the same feature extraction network is proposed. First, ResNet-50 network is used to extract image features of the encoder. Then, multi-scale feature prediction and fast regression in single shot multibox detector target detection algorithm are used to decode the detection results. A pyramid pool structure of porous space in DeepLab v3 is used to process the multi-scale mapping, bilinear sampling and batch normalization of the image features after ResNet-50 sampling so as to complete segmentation and decoding. Finally, the training of the joint method is completed under the set training parameters. Experimental results show that the mean average precision of the method is 89.00%,the mean intersection over union is 83.0, and the number of frames per second is 31 frame, which can support intelligent vehicle to complete certain tasks.
李立凯, 卢炽华, 邹斌. 基于深度学习的目标检测与可行域分割研究[J]. 激光与光电子学进展, 2020, 57(12): 121013. Likai Li, Chihua Lu, Bin Zou. Research on Target Detection and Feasible Region Segmentation Based on Deep Learning[J]. Laser & Optoelectronics Progress, 2020, 57(12): 121013.