Image Pattern Learning And Application Based On Deconvolution Network

Deep learning has been appiled in various fields of study and already made remarkable progress rapidly. Convolution network, one kind of deep learning, performing well both in computer vision and pattern recognition, has been proven a potential and promising work in the furture. However these works are mainly focus on the study of high-level semantic model of nerual network. Instead, the main work of this paper is to extract given images' internal features, like color, texture and spatial distribution, based on deconvolution network.The main goal of this paper is achieved through the following three aspects: firstly, the design of the network structure is a deconvolution network, which means it not only have convolutional part based on VGG 16-layer net, but also deconvolutional part that is a mirrored version of the convolution network, having unpooling and deconvolution layers. Such a design can capture details that may lose during the covolution, and we can also customize learning even in the same network structure by adjusting initial weights in fullconnection and normalization layer to response different characteristic patterns.Secondly, to get a stable and multi-scale representation of image feature in the design of function model, we use Gram matrix[13]to calculate correlation between different filter responses of different scale kernels in every layer, and build up a loss function which is a linear summation of correlation of filter responses.Lastly, we would also consider objects' spatial relationship in image. When the feature we need to learn is just a significant region on the image, not the whole image, we propose a method to extract the region contour, and then learning would just happen in this constraint region.In addition, we describe some applications based on our learning model, like texture synthesis and image style transfer. Artistic style and image style based on spatial relationship would be discussed seperately in the style transfer part. On the experimental comparison, we would compare our result on texture synthesis with sample-based synthesis[42] and parametric synthesis[51]. Then we would compare our result on image style transfer with Gatys' [11] work, which use convolution network for style transfer.