摘要:在中国,智能手机的高普适性增加了用户对于图像的需求量。网上的图像资源也随之快速增长,如何让用户快速获得所需的图片成为了目前的一大难题。一种可靠的解决方案是加深图像的归类,即在图像归类时,给图像打上接近于人类自然语言的高级语义标签。但是机械对于图像视觉特征只是单纯地记录(底层图像语义),而人类对于图像视觉特征进行了自我的理解与演绎(高层图像语义),这就造成了语义鸿沟问题。目前通过深度神经网络的使用,高级语义的提取问题已经得到很好的解决。另一种可行的解决方案是根据用户的需求生成图像。要实现这一点,首先要让机器变得更加智能,学习如何根据图像语义自行生成图像。因此,本文研究了卷积神经网络和生成式对抗网络的相关深度学习理论,同时利用实验研究图像生成的问题。80882
毕业论文关键词:深度学习;图像语义;图像生成;生成式对抗网络
Image Semantic Extraction Based on Deep Learning
Abstract: In China, the high popularity of smart phones increases the user's demand for images。 Online image resources also grow rapidly。 How to make users to quickly get the required images has become a major problem。 A reliable solution is to deepen the classification of images, that is, when the images are classified, the images are marked with high-level semantic labels which are close to the human natural language。 But the image visual features are simply recorded by machines (the underlying image semantics), yet human has deep understanding and interpretation of image visual features with their own feelings and knowledge (high-level image semantics), which causes semantic gap。 At present, the problem of advanced semantic extraction has been solved very well through the use of Deep Neural Networks。 Another viable solution is to generate images according to the needs of the users。 To achieve this, the machine need be more intelligent first, and learn how to generate images on their own with image semantics。 Therefore, this paper studies the related deep learning theories of Convolutional Neural Networks and Generative Adversarial Networks, then studies image generation with experiments。
Key Words: Deep Learning; Image Semantics; Image Generation; Generative Adversarial Networks
目录
一、绪论 1
(一) 背景 1
(三) 研究目的与意义 4
二、深度学习相关理论研究 6
(一) 历史 6
(二) 深度学习原理 7
(三) 卷积神经网络 9
(四) 生成式对抗网络 11
(五) 辅助分类器生成式对抗网络 13
(六) Wasserstein生成式对抗网络 13
(七) 梯度惩罚Wasserstein生成式对抗网络 15
三、WAGN图像生成实验研究 17
(一) 实验目的与内容 17
(二) Keras 17
(三) 实验环境 18
(四) 图像数据集 18
(五) 简易WGAN实验 18
(六) AC-WGAN实验 19
(七) WGAN-GP实验 21
(八) 结论