The Research On Multiple Deep Neural Networks Generative Adversarial Mechanism For Image Cognition

Generative Adversaria Nets(GANs)which consists of multiple deep neural networks aim to fit the real data distribution by learning the nonlinear mapping between initial and real distribution,and further generate various types of data including image,audio,and video.The advantage of generative adversarial nets is that it removes the Markov constraints,it has strong compatibility with other models and it requires no complex reason.This dissertation focuses on the research of multiple deep neural networks generative adversarial mechanism oriented to image cognition.Considering the research hotspots of generative adversarial nets,this dissertation concentrates on the aspects of image restoration,hash retrieval and cross-domain prediction as well as theoretical research and application development.The main contributions can be summarized as:1.To solve the problem of vanishing gradient,unstable training,poor local consistency and time-consuming in GANs-based image restoration,multi-granularity generative adversarial nets with reconstructive sampling is proposed for regular and irregular pattern inpainting.The method firstly modifies the low-dimension noise sampling to reconstructive sampling for non-zero Lebesgue measurement,which stabilizes the gradient and training with theoretical guarantee.Besides,the segmentation invariance is introduced to reduce the training time and parameters on the granutee of satisfactory inpainting results,and the algoritm admission is proposed to comprehensively measure inpainting methods.The experimental results show that the proposed method can achieve satisfactory inpainting results and acceptable training time.2.Considering that the current image hash retrieve algorithms are only suitable for complete data retrieval,an incomplete data hash retrieval algorithm based on multigenerative adversarial nets is proposed.The proposed algorithm integrates hash,restoration,discriminative and classification network into an end-to-end framework and realizes effective retrieval of incomplete data for the first time.In addition,a new similarity maintance measurement,supervised manifold similarity,is proposed to preserve similarity among similar samples and dissimilarity among dissimilar samples.Theoretical demonstration is provided to prove that it has superiority compared with point-wise or pairwise similarity metric.Experiments show that proposed method not only obtains absolute increment in incomplete data retrieval,but also shows competitiveness in complete data retrieval due to the introduction of supervised manifold similarity.3.Considering that current hashing algorithm only achieves the goal of fast retrieval but ignore the storage advantage of hash code.This study proposes a bidirectional transformation method between hash code and image based on multi-generative adversarial nets,which is the first attempt towards this target in the information retrieval research community.The proposed method combines the hash coding and image synthesis,and uses supervised manifold similarity to improve the retrieved accuracy and user's acceptance.Experiments show that the conversion of large-scale data into hash codes through the hash network greatly promotes the reduction of storage space while ensuring high retrieval accuracy and good user's acceptance.Then,hash code can be used to reconstruct into these close to original images by using the inverse hash network.4.Aiming at the problem that there is current no method which realizes bidirectional cross-domain prediction in medical imaging,this work presents a novel method for bidirectional prediction between CT and MRI image based on multi-generative multiadversarial nets.In order to ensure the homeomorphism mapping between source and target domain,the localization is exposed on predictor(generator)to achieve local prediction,where auxiliary label information is introduced to constrain feature generation and attack the potential risk of pathological variance,and the edge retention metric is proposed to preserve the anatomical structure.Then,spectral normalization is proposed to control the performance of discriminatior and accelerate learning speed and improve predicted quality.The extensive experiments show that the proposed can yield satisfactory predictive results when obtain CT or MRI images.Besides,the experimental results of paired and unpaired MRI-CT show that predicted images by proposed algorithm have obtained state-of-the-art evaluations subjectively and objectively.