Research On Visualizing Image Features Based On Convolutional Neural Networks

Convolutional neural network,as an important model in the field of deep learning,has achieved great success in most fields of computer vision.However,the interior of a deep neural network is like a "black box",and its working principles are never understood.Therefore,to better understand the inner operating principles of convolutional neural network,deep visualizing has been developed as a new research field with the purpose of analyzing image features extracted by convolution neural network.Deep visualizing can help understand how each layer of convolutional neural network extract features and which information of an image is preserved,which is useful to avoid blind parameter finetune and trial-and-error in the network training process,thus quickly achieving optimal performance of the network architecture.One work of this article is to make visualizing analyze to the inner structure of Convolutional Neural Network.Due to the fact that each kernel extracts a particular feature map,and the kernel in lower layer extracts the texture information,while the features extracted by the higher-level convolution layer are the combination of the upper layers,thus forming more abstract target features.Based on this conclusion of the image features visualization,this paper proposes a style transfer method based on image iteration,which can further generate a new image with transferred style.The main contributions of this paper are summarized as follows:(1)The deconvolution technique is used to visualize feature maps extracted from each layer of convolutional neural network.The visualizing results show that the lower convolutional layer mainly extract the image edge,texture and color information,and the high-level layer of neural network will extracte more abstract feature maps.(2)A visualizing method based on class activation map is proposed to visualize and analyze each layer of the convolution neural network.The experimental results show that this method can effectively illustrate the important area of an image in tasks such as classification especially in the last convolutional layer,and make a reasonable explanation about why convolutional neural network can achieve such a good effect in the classification task.(3)According to the research results of deep visualization,namely lower convolutional layer extract the texture feature,while high-level convolutional layer extract relatively abstract content information,the initial random noise image can be iteratively changed by minimizing the total loss function of content and style loss,making the iterated image retained both content information and style texture information,ultimately achieving the effect of style transfer.