Research On Effective Receptive Field Of Deep Convolution Neural Networks

In recent years,deep learning based on deep convolutional neural network has developed rapidly and achieved a leading position over traditional algorithms in the fields of computer vision.Convolutional neural network is an important technology in deep learning.However,although the deep convolutional neural networks perform well in many fields,human can't understand how the networks make decisions,deep convolutional neural networks based deep learning technology is considered as "black box" with its credibility being questioned.Therefore,the interpretability of deep convolutional neural networks is an important research fields.Effective receptive field refers to the area of input pixels in the convolutional neural network that has a non-negligible influence on the output unit.effective receptive field can be used to indicating how many contributions that input pixels make to decision.Thus,it provides an explanation for the decision of deep convolutional neural network.This dissertation focus on the effective receptive field of deep convolutional neural networks,and analyzes the effective receptive field of the network from both experimental and theoretical aspects.In order to construct the effective receptive field algorithm,the vector form of the gradient in the back propagation process is derived in this dissertation firstly.The relationship of the forward propagation step size and the back propagation calculation is obtained.According to the derivation,the effective receptive fields of deep pure convolutional networks and convolutional networks with different activation functions are constructed.The results show that the effective receptive fields are Gaussian and the activation function has a great influence on the effective receptive fields.The experiment result is proved using Fourier transformation.At the same time,the analysis of the effective receptive field of the activation function,under the assumption that the input is a lognormal distribution,draws the same conclusion.To explore the relationship of effective receptive fields and the performance of network,the effective receptive fields of four typical deep convolutional neural networks are studies and analyzes in this dissertation.First,the internal calculation form of AlexNet,VGG16,YOLOv1,ResNet18 network are analyzed and derived;then according to the derivation result,the effective receptive fields of the corresponding networks are constructed.Experimental results show that the effective receptive fields of these networks are all Gaussian.At the same time,according to the statistical analysis of experimental results,the depth of the network is positive correlated with the area of the effective receptive field.Moreover,the network performances better as the effective receptive field area rate being higher.