摘要
针对Faster-RCNN主干特征提取网络Resnet50原本的ReLU函数性能仍有提升空间这样的问题。论文探索使用高斯误差线性单元激活函数GELU去提升原网络的性能。高斯误差线性单元激活函数GELU相较于ReLU激活函数,用概率的方法去保留有效信息,而不是ReLU函数那样单纯的对于大于0的信息保留,小于0的信息舍弃。这使得用GELU作为激活函数能更准确反映图像信息,从而提高网络整体的目标检测性能。实验表明,改进后的网络的mAP相较于原网络有了一定的提升,在识别道路信息时有更好的表现,能够更好地应用于无人驾驶技术。
There is still room to improve the performance of the original ReLU function of the Faster-RCNN backbone feature extraction network Resnet50.The use of Gaussian error linear element activation function GELU is explored to improve the performance of the original network.Compared with the relu activation function,the Gaussian error linear unit activation function GELU uses the probability method to retain the effective information,rather than the ReLU function,which simply retains the information greater than 0 and discards the information less than 0.This makes using GELU as the activation function can more accurately reflect the image information,so as to improve the overall target detection performance of the network.Experiments show that compared with the original network,the map of the improved network has a certain improvement,has better performance in identifying road information,and can be better applied to driverless technology.
作者
张起航
曾琦
ZHANG Qihang;ZENG Qi(Tianjin University of Technology,Tianjin 300384)
出处
《计算机与数字工程》
2024年第8期2411-2415,2419,共6页
Computer & Digital Engineering
作者简介
张起航,男,硕士研究生,研究方向:深度学习图像处理;曾琦,女,硕士研究生,研究方向:深度学习图像处理。
