全卷积神经网络仿真与迁移学习

人们捕获视图,从视图中提取特征并理解含义。同理,驾驶员也通过视觉实现对街景的判断。我们期待,有一天机器能够通过自主计算完成同样的工作。得益于计算机的强大处理能力,基于CNNs(Convolutional Neural Networks,卷积神经网络)的深度学习算法能够很好地完成目标识别等计算机视觉任务。但在实际工业应用中,资源往往受限,较大的网络不利于嵌入式移植。通常一个完备的CNN网络包含卷积层、池化层和全连接层1],本文参考文章2]中的方法,舍去池化层和全连接层,使用卷积层代替,并对几种网络进行了仿真实验及结果分析,寻找在受限平台使用CNN网络的方法。

All Convolution Neural Network Simulation and Transfer Learning

As we know,vision is very important for human to understand the world.Similarly,drivers also use vision to anlysis street scene.What we had being expecting is that one day machine can do the same work by itself through computer vision.Thanks to the powerful processing ability of the computer,the deep learning algorithm based on CNNs (Convolutional Neural Networks) can accomplish computer vision tasks well such as object detection. However,in most industrial environment,resources are limited,and larger networks are not easy to embedded transplantation.Now,the architectures of main CNNs used for machine vision include convolution layers and pooling layers alternately and ended with several fully-connected layers 1].In fact,we can use convolution layers to replace pooling layers,as described in paper 2].