Key Technology Research And Application Of Intelligent Police Sketch System

Police Sketch are drawn by portraitists through the descriptions of witnesses,combined with criminal physiognomy and the personal experience of the painters,to paint portraits of suspects,which has the most important role in criminal investigation work.Due to its professionalism and complexity limitations,the threshold of use is high.The existing Police Sketch System mainly relies on image stitching and synthesis technology.Which has the problems of raw stitched face images,lack of enlightening recommendations,inflexible face editing,missing details of portrait faces and not easy for mass retrieval.This thesis focuses on the problems existing in the analog profiling system.The specific research content is as follows:(1)At present,image translation mainly relies on the proposed mapping of data existing in pairs,but more data of portrait scenes in Police Sketch are not in pairs.Therefore,the pairwise generation network in the unpaired Cycle GAN is improved based on the directional condition and the design of parameter sharing for the Police Sketch characteristics.The experimental results show that the network proposed in this thesis,compared with Cycle GAN,reduces network parameters by 34%,improves the PSNR value on generated images by 6.25% and SSIM value by 17.65%,and significantly improves the quality of reconstructed images.(2)The lack of enlightening recommendations when conducting simulated portraits in the current simulated portrait system leads to a high threshold of system use and uncoordinated portrait results in face components;face attribute editing relies on dataset tags,and the main methods and data target color faces,which are not applicable to portrait face editing for criminal investigation simulated portraits.Therefore,this thesis performs intelligent classification and colloquial mapping of face components,and face recommendation algorithm based on association mining to achieve face recommendation under only partial clues.The proposed style coding based on color face data on the extraction of editing direction solves the problem of portrait face editing,and tuning through the key point change law realizes a more convenient and flexible portrait face editing network to meet the complex requirements under criminal investigation scenarios.(3)In Police Sketch Systems,which is not applicable to the existing face recognition network retrieval,so the color reconstruction is needed.The color face photos generated by GAN often suffer from blurring and detail loss due to the limitation of computational resources and data sets.Current image super-resolution methods rely on fixed degraded data for supervised learning,which is not applicable to degraded images containing complex noise in real scenes.Therefore,this thesis proposes a self-supervised approach based on a self-supervised approach to compute the loss by oversampling and resampling on the potential space of the generative adversarial network.Through experimental verification,the proposed method can generate high-resolution realistic images on edited deteriorating color reconstructed faces,effectively improving face clarity and increasing recognition.The Naturalness Image Quality Evaluator of face images improves 23.6%.