Method Research On Person Image Generation Based On Deep Learning

With the development of technology,people's demand for security monitoring is rising rapidly.Therefore,many cameras placed in key buildings and densely populated areas constitute a camera network for real-time monitoring.These security monitoring systems provide people with security.But processing a large amount of data increases the people burden.In order to automatically figure out the contents in the videos,person re-identification(re-id)has received extensive attention among the scholars.At present,although scholars from various countries have proposed many person re-id algorithms on the basis of deep learning,compared with pattern recognition methods and machine learning methods,has made significant progress,but person re-id still faces two major problems.The first problem is the datasets used to train the neural network are insufficient in scale and content.The second problem is the misalignment between samples,including background,illumination,pose,clothing,etc..And the misalignment of pose has great impact on the performance of the models.Therefore,the main research contents of this thesis are two aspects,which are studying how to convert the source pose image to the target pose on the basis of GANs,and generating the target pose image with the specified pose feature.This paper first studies how to convert input pose to target pose.Based on the Pose Guided Person Image Generation method,this paper further uses pose mask map and key points binary-heat map to guide the pose transfer network and multi-scale differential refine network.This paper introduces identity category as prior knowledge to supervise network.Based on the residual module and the skip connection method,this paper further proposes the multi-scale residual module and the adaptive skip connection method,so that the network can not only learn multi-scale features from images,but also improve the efficiency of feature transmission,as well as reducing feature loss.In terms of network training,this paper proposes the neighborhood loss and the smooth loss to reduce the spatial difference and the high-frequency noise.The experiments prove the rationality and effectiveness of these proposed methods,and achieve higher pose accuracy conversion results.In order to avoid the Background interference from the source input image,this paper studies how to reconstruct the complete person image according to the specified target pose information.In order to ensure the feature description meet the requirements of reconstructing,this paper proposes the concept and acquisition method of the conditional pose person feature,which is used for reconstructing the complete person image by the residual reconstruction network and repair network.A large number of experiments prove the effectiveness and rationality of the conditional pose person feature and reconstruct more accurate person images.At the same time,this paper studies the influence of network parameters and scale on the the reconstruction network.Based on the weight-pruning algorithms,the conditional pose person feature is masked by a trainable layer.This make the reconstruction network can be trained twice.We prove that soft mask and hard mask can optimize the internal feature representation by compressing parameters,which provides a valuable research perspective for the interpretability and compressibility of neural networks.