基于全连接神经网络的近实时重新着色算法  

Near Real-time Image Recoloring Method Based on Fully Connected Neural Network

作　　者：厉旭杰[1] 王怡婷 LI Xujie;WANG Yiting(College of Computer Science and Artificial Intelligence,Wenzhou University,Wenzhou 325035,Zhejiang,China)

机构地区：[1]温州大学计算机与人工智能学院,浙江温州325035

出　　处：《实验室研究与探索》2020年第2期21-24,共4页Research and Exploration In Laboratory

基　　金：浙江省自然科学基金项目(LY18F020022)。

摘　　要：提出了一种基于全连接神经网络(FNN)的图像重新着色算法。该算法提取着色线条所在区域的像素RGB颜色特征值和相应的着色线条分类为数据集,为了减少神经网络的训练时间,对数据集进行了采样;把FNN作为一个像素级的多分类神经网络,利用这些训练数据训练FNN,将待重新着色图像中逐个像素的特征值作为神经网络的输入,获得每个像素属于着色线条的似然概率;根据神经网络输出的每个像素属于着色线条的似然概率,计算最终的图像重新着色结果。与现有的基于卷积神经网络的图像重新着色方法相比,该方法避免了神经网络在训练阶段需要大规模的训练样本的弊端,且能够达到近实时的交互性能,同时用户只需输入少量的用户着色线条,就能获得高质量的图像重新着色效果。A simple but very effective fully connected neural network model (FNN) on the image recoloring is proposed. This method extracts the feature vector in the region where the user strokes are located and the corresponding classification of user strokes into dataset. To reduce the whole training time,an efficient sampling strategy is proposed.The fully connected neural network model on the image recoloring is regard as a multi-class pixel-level classification problem. The extracted dataset is fed as input to train our FNN-based model. Then,each feature vector from the whole image is fed as input. The trained FNN-based model is used as a classifier to estimate the probabilities. Finally,the image recoloring result is calculated using the probabilities. In contrast to previous image-level CNN-based network,the proposed method avoids demanding a large number of images in the training stage and achieves the near real-time processing for the image recoloring using the user strokes. Meanwhile,our approach produces high-quality results with only a small amount of user interaction.

关 键 词：全连接神经网络 图像重新着色 特征值 像素级分类神经网络 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]