基于卷积神经网络的图像编辑传播

摘要

For the problem of editing propagation of a single image, a combinational convolution is introduced instead of traditional convolution to extract more effective features.Combinational convolution is composed of the deformable convolution and the separable convolution.This structure can enhance the generalization ability of the model.Besides it can reduce the parameters of the model and the operations of the convolution.At the same time, it introduces a weighted loss function for the background class to prevent the pixels which are have the high degree similarity of the background from being erroneously colored and causing color overflow.The experimental results show that the bi-branch convolutional neural network model constructed by combinational convolution and the biased loss functions can effectively colorize a single image and improve color overflow.%对于单幅图像进行编辑传播的问题, 引入组合卷积来代替传统的卷积, 用以提取更加有效的特征.组合卷积由可变形卷积和可分离卷积组成, 通过这个结构可以增强模型的泛化能力, 并且减少模型的参数量和卷积的操作数.同时引入对错分的背景类进行加权的有偏损失函数, 以防止与背景类相似度较高的像素点被误着色而造成颜色溢出.实验结果表明:使用组合卷积和有偏损失函数构建的双分支的卷积神经网络模型, 可以实现单幅图像的有效上色, 并且能够改善颜色溢出的情况.