Application Of Deep Learning In Image Processing

This paper applies deep convolutional neural network for image editing propagation and image style transfer.For the problem of the single image editing propagation,first we introduce a combinational convolution instead of the traditional convolution to extract more effective feature maps.Combinational convolution consists of the deformable convolution and the separable convolution.It can extract more reasonable visual features and reduce the model parameter amount and convolutional operations by using this combinational convolution.At the same time,we introduce a weighted loss function for the background class of the misclassification to prevent the background class from being erroneously colored and causing color overflow.The experimental results show that the proposed model can effectively paint single images and improve color overflow.For the problem of the image style transfer,we build a style transfer convolutional neural network based on the combinational convolution.It uses the combinational convolution to extract the content features of the content image and the stylistic features of the style image.Then we construct the loss function based these features and obtains the stylizedimage by the constraint of the loss function.The model uses the adaptive instance normalization layer which make a single neural network model produce any stylized image possible.Experimental results show that this method can quickly obtain stylized images,and the generated images can retain a strong style.