Research On Collaborative Generation Guidance Algorithm For Visual Analysis

When dealing with visual analysis problem,general deep learning methods try to learn the function mapping from input space to solution space,ignoring the task specific prior knowledge.Therefore,most deep learning methods are limited by training data,which makes them lack of generalization ability and poor of interpretability.In order to utilize the training data and prior effectively,this paper proposes a flexible collaborative generation guidance algorithm framework.Specific algorithms for different visual tasks can be formed by designing the data-driven generation module and task prior guidance module.This paper focuses on solving blind image deblurring problem.The edge structure is very import to estimate the blur kernel.This paper designs the generation module as a convolution residual network with initialization layer,which is used to generate potential clear image with significant edge from degrade image.In the process of inference,network lacks mathematical guarantee,so its output is easy to deviate from the predetermined optimization direction.Therefore,the guidance module is designed as a corrector based on energy discrimination and prior proximal operator,to correct the output of network.The corrector ensures that the potential image sequence meets the monotonicity of object energy function and its first-order optimal necessary conditions,making the whole solution process is theoretical convergence guaranteed.By modifying the generator and corrector,the algorithm can be used to solve other low-level image processing problems.The proposed algorithm framework is also applied to skin segmentation task.This paper designs a new dual task network with mutual guidance,in which the generation module can output the skin mask and body mask simultaneously.The body mask is used as prior information of skin area to further refine the skin segmentation result in the guidance module,and the skin mask is also used as prior information of body area.By using cheap body mask,this paper can get better segmentation results on the limited skin labeling data set.In addition,this paper also propose a weighted semi-supervised loss function with label existence index,which contains the conditional random field segmentation prior and body-skin belonging prior.By using this newly designed loss function,dual task network can be trained on incomplete training data and maintain good performance.