Neighborhood Selection Optimization Based Heterogeneous Image Transformation

Video surveillance face recognition can help police quickly identify and track target persons.With its unique application advantages in public security,video sruveillance face recognition becomes a hot issue in recent years.In practice,affected by the factors such as shooting angle,shooting environment and so on,heterogeneous image transformation methods like face image super-resolution and face photo-sketch synthesis are often adopted to trasform the target face image in video into high resolution,frontal face image in order to improve the following recognition rate of face recognition.In this paper,we use face sketch generation as example to explain the existing example-based heterogeneous image transformation method.In view of the shortcomings of the current methods,we make method innovation in the neighborhood selection stage and optimize candidate neighborhood,to enhance the quality of synthesized image.The main innovations of this paper can be summarized as follows:1.We proposed an anchored neighborhood index based face sketch synthesis method.Most of current example-based face sketch synthesis methods are based on the assumption that the photo-sketch patches are locally isometric,thus only photo patches are utilized to search candidate neighborhood and optimize weights,which neglects the sketch manifold,causing low quality such as blurring and noise in syntheized sketch.In view of this situation,we propose an anchored neighborhood index based face sketch synthesis method,which makes full use of the co-occurrence priors of the corresponding photo-sketch patches in the training set.The selected candidate photo-sketch patches are similar simultaneously,with a higher degree of local isometry.Experimental results demonstrate that the proposed method can reduce the noise of the candidate image,and greatly improve the image definition.2.We proposed a Bayesian inference based face sketch synthesis method.Aiming at the situation that current example-based face sketch synthesis method easily selects candidate image patches which are deviated from the actual situation,spatial neighborhood smoothing constraint is added into neighborhood selected stage,thus we proposed a Bayesian inference inference based face sketch synthesis method.The method first use traditional approaches as initial neighborhood selection,then the selected initial candidate neighborhood is brought into markov random field model,which further optimize the reconstruction and use the spatial neighborhood smoothing "optimal" top K pairs of photo-sketch patch as final candidate image patches.Finally markov weight field is utilized to weight optimization and the following sketch synthesis.Experimental results show that the proposed method can choose candidate image patches that are more suitable to the real situation,which makes the synthesized image with higher fidelity.3.We proposed a neighborhood selection optimization based face sketch synthesis method.This method combines the advantage of the anchored neighbor index based and bayesian inference based neighborhood selection method,which firstly selects initial candidated neighborhood using the anchored neighbor index based method from the training set.The markov random field is used to seletct a pair of photo-sketch patch with the optimal reconstruction and spatial neighborhood smoothing,thus obtain the coupling K nearest neighbor of the photo-sketch patch.Finally markov weight field is utilized to weight optimization and the following sketch synthesis.Experimental results show that the proposed method can choose candidate image patches that are with a higher degree of local isometry and more suitable to the real situation,which makes the synthesized image with higher definition and image fidelity.