【中文】一种基于神经网络的多风格国画花生成方法 【EN】Multi
摘要【中文】本发明提供了一种基于神经网络的多风格国画花生成方法，本发明使用的注意力架构学习从白描到工笔或水墨的映射，并使用非成对的数据集训练，得到了具有一定风格特点的工笔和水墨画；本发明引入一个新的损失MSSSIM损失函数在基于GAN的注意力架构中，从亮度，对比度，结构三个方面多尺度地比较重构图像与源图像的距离，并与循环一致性损失L1函数共同引导图像风格和内容的准确生成，解决了原网络画面不清晰，出现伪影，墨迹模糊一片，存在噪点等缺陷的问题。本发明解决了在传统方法的研究上，由于国画风格的抽象性和复杂性，不能够生成画面清晰地高质量图片的问题。 【EN】The invention provides a multi-style Chinese painting generation method based on a neural network.A attention architecture used by the method learns the mapping from white drawing to a pen or ink, and uses unpaired data set training to obtain the pen and ink with certain style characteristics; the invention introduces a new loss MSSSIM loss function in an attention architecture based on GAN, compares the distance between a reconstructed image and a source image in a multi-scale mode from three aspects of brightness, contrast and structure, and guides the accurate generation of the style and content of the image together with a cycle consistency loss L1 function, thereby solving the problems of unclear original network picture, artifact, fuzzy ink, noise and the like. The invention solves the problem that clear high-quality pictures can not be generated due to the abstraction and complexity of the style of traditional Chinese painting on the research of the traditional method.
相关知识
一种基于多传感器信息融合的故障诊断方法
基于卷积神经网络的多尺度葡萄图像识别方法
基于BP神经网络的农作物虫害预测系统的研究
基于神经网络的食用玫瑰花图像识别算法
基于卷积神经网络的花卉识别方法
基于深度卷积神经网络的植物病虫害识别
基于BP神经网络的棉花病虫害诊断系统
一种基于LVQ神经网络的黄瓜温室温度智能化检测装置制造方法及图纸
基于卷积神经网络的农作物病虫害识别系统
基于超轻量级全卷积神经网络的花卉识别方法
网址: 【中文】一种基于神经网络的多风格国画花生成方法 【EN】Multi https://m.huajiangbk.com/newsview1582708.html
