摘要
为了解决传统的卷积神经网络着色方法带来语境混淆、边缘模糊和细节信息丢失度高等问题,设计了一个改进的基于密集神经网络的区域全卷积神经网络(R-FCN)和基于局部特征网络的双分支神经网络模型.通过DenseNet可准确提取信息,产生易于训练和高参数效率的密集模型,采用全连接的条件随机场优化分割结果来提高分割的准确率.验证阶段采用联合双边滤波对图像进行处理,弥补图像边缘模糊的缺点.实验结果表明:与现有着色方法相比,该方法有效地解决了细节丢失度高、颜色不饱和及边缘模糊的问题,能够产生更真实、更合理的彩色图像,取得了优异的效果.
In this study, we design an improved regional full convolutional neural network (R-FCN) based on dense neural networks and a two-branch neural network model based on the local feature network to solve the problems of context confusion, edge blur, and high loss of detail information associated with the traditional convolutional neural network coloring method. DenseNet is used to accurately extract information and create a dense model, which can be easily trained and exhibits high parameter efficiency, and fully connected conditional random fields is used to optimize the segmentation results for improving the segmentation accuracy. During the verification stage, joint bilateral filtering is used to process the image to solve the disadvantages associated with image edge blurring. Experiments results show that, compared with existing coloring methods, the proposed method can effectively solve the problems of high detail loss, color unsaturation, and edge blur, produce more realistic and reasonable color images, and achieve excellent results.
作者
何山
方利
张政
He Shan;Fang Li;Zhang Zheng(School of Com puter Science,Southwest Petroleum University,Chengdu,Sichuan 610500,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2020年第12期117-123,共7页
Laser & Optoelectronics Progress
关键词
图像处理
语境混淆
密集神经网络
区域全卷积神经网络
全连接条件随机场
联合双边滤波
image processing
context confusion
dense neural network
regional full convolutional neural network
fully connected conditional random field
joint bilateral filtering
