【作者】 张晶；

【导师】 顾正弘； 周平；

【作者基本信息】 扬州大学 ， 工程硕士（专业学位）， 2021， 硕士

【摘要】 一直以来,图像处理是一个具有难度和挑战的问题,并且对于图像处理中发生器模型的建模也极为不简单,一般由于后验概率的取值极值化进行优化求解,在现如今大数据的时代,信息爆炸使得求解更为艰难。随着深度神经网络的不断研究突破,在图像识别、语音识别以及自然语言处理方面有较为不错的应用能力,并且在准确性方面也有着显著的提升。但由于更深层次的研究,诞生了一种生成对抗神经网络模型,此模型为对抗性学习开辟了一种新格局。生成对抗神经网络对比传统神经网络模型有着显著的差异,其包含两种神经网络模型结构,一个是生成器,而另一个是判别器:其中生成器模型主要负责生成样本,判别器则主要负责判别样本的真假,生成器网络模型的工作原理是输入随机噪声进入其神经网络之后试图生成新的样本数据,而判别器工作原理则是将真实数据或生成的数据输入到其网络模型中进行学习,并且判断生成器生成的数据是否为原始数据,生成器输入的样本数据从原先真实的数据分布Pdata(x)中获取,通过判别器判别之后,接下来就是解决一个二分类问题,在判别真实性之后判别器产生0-1之间的数值来判别输入的样本真实性的概率。然而这个模型也存在一些问题:在训练模型时会出现梯度爆炸、梯度消失以及生成图片要求达不到预期等问题,本文提出了一种改进后新的生成对抗神经网络模型,通过引入特征标签引导学习并且在判别器和生成器中加入均方误差损失,使得生成指定的样本,本文的主要工作以及研究成果如下:1.为了进一步提高生成对抗式神经网络(GAN)[1]的准确性,针对该算法模型进行的改进工作有:(1)优化GAN神经网络结构,在其原先生成的随机性图片的基础上引入了条件的概念;为了更好地生成图片,在GAN网络中对于生成器和判别器添加了约束条件得以实现指导作用,用来解决GAN网络训练自由度较大的问题。(2)考虑到GAN网络的稳定较低,在训练的过程中容易发生梯度爆炸、梯度消失以及训练不平衡等问题,在建立改进模型中,将判别器结构改为自编码器,并提出边界平衡的理念。在原先的基础上再次引入边界平衡的算法[2],提出了一种超参数k来作为平衡依据,目的是通过这个超参数设定可以在图像的复杂性和图像的生成质量之间做均衡。2.基于条件生成对抗神经网络(CGAN)的模型原理,改进了CGAN模型的算法,使得其稳定性更加优秀,改进后的模型有效解决了原始GAN模型的缺乏引导性以及稳定性差,提供了更加高效且稳定的有监督学习算法,在一定程度上提升了图像处理的效率和质量。更多还原

【Abstract】 The image processing problem is a difficult and challenging problem.It has always been extremely difficult to model the image generation model.Generally,it is optimized to solve the problem by maximizing the posterior probability.In the era of big data,information The explosion makes the solution more difficult.With the continuous research and breakthroughs of deep neural networks,they have relatively good application capabilities in image recognition,speech recognition,and natural language processing,and they have also significantly improved their accuracy.However,due to deeper research,a generative adversarial neural network model was born,which opened up a new pattern for adversarial learning.Generative adversarial neural networks are significantly different from traditional neural network models.Its composition is divided into two parts of the network model,namely the generator and the discriminator:the generator is responsible for generating samples,the discriminator is responsible for identifying the true and false of the samples,the generator The working principle of the network model is to input random noise into its neural network and try to generate new sample data,while the working principle of the discriminator is to input real data or generated data into its network model for learning,and then determine whether the data generated by the generator is Real data,the sample data input by the generator is obtained from the original real data distribution.After being discriminated by the discriminator,the next step is to solve a binary classification problem.After discriminating the authenticity,the discriminator generates a value between 0-1 to distinguish The probability of the authenticity of the input sample.However,this model also has some problems:gradient explosion or gradient disappearance in the training model,and the generated images cannot meet our conditions.This paper proposes a generative adversarial neural network model with conditional boundary balance,which guides learning by introducing feature tags And the mean square error loss is added to the discriminator and generator to generate the specified sample.The main work and research results of this paper are as follows:1.In order to further improve the accuracy of the Generative Adversarial Neural Network(GAN)[1],improvements to the algorithm model are as follows:(1)Optimize the GAN neural network structure based on the original random images generated The concept of conditions is introduced;in order to better generate images,constraints are added to the generator and discriminator in the GAN network to achieve a guiding role,which is used to solve the problem of greater freedom in GAN network training.(2)Considering the low stability of the GAN network,the gradient explosion,gradient disappearance,and training imbalance are likely to occur during the training process.In the establishment of the improved model,the discriminator structure is changed to the autoencoder,and the boundary is proposed Concept of balance.On the original basis,the algorithm of boundary balance is introduced again.By providing a hyperparameter k,this hyperparameter can balance the diversity of the image and the quality of the generated image[2].2.Based on the model principle of Conditional Generative Adversarial Neural Network(CGAN),the algorithm of the CGAN model is improved to make its stability more excellent.The improved model effectively solves the lack of guidance and poor stability of the original GAN model,It also provides a more efficient and stable supervised learning algorithm,which improves the efficiency and quality of image processing to a certain extent.更多还原