基于图像云模型语义标注的条件生成对抗网络

Conditional Generative Adversarial Network Based on Image Semantic Annotation of Cloud Model

在图像补全技术中,当图像丢失较多信息时,仅凭自身已有的信息很难补全图像.因此,文中使用条件生成对抗网络(CGAN)和多粒度认知相结合的方式研究图像的降噪和补全.首先借助云模型中高斯云变换算法提取无标签图像的多层语义信息,并根据不同层次的语义信息对图像进行不同粒度的分割,同时对已分割图像进行自动语义标注.然后将各粒层图像和其对应的语义信息分别作为CGAN的训练数据,得到图像生成对抗网络模型.最后依据此模型补全图像的缺失信息.实验表明,对于Caltech-UCSD Birds和Oxford-102flowers数据集的图像降噪和图像补全,文中算法取得较好效果.

As the missing information in the image is increasing,the existing methods extracting information from only a single image can not produce satisfactory completion results. Therefore,an automatic label conditional generative adversarial network（ CGAN） based on image semantic is presented from the perspective of multi-granular cognition. It can be applied on image denoising and image completion. Firstly,the multi-layer semantic information from unlabeled images based on the Gaussian cloud transform algorithm is extracted. Then,the original images are segmented and the segmented images are labeled automatically in accordance with different granular semantic information. Furthermore,different granular segmented images and their labels are used as the training samples in the CGAN to get an image probability generation model,respectively. The large missing regions from a single image are completed based on the similar image generated by cloud semantic and CGAN. On the datasets of CaltechUCSD Birds and Oxford-102 flowers, the proposed model achieves the high performance in image denoising and image completion.