Research On Mobile Spectral Video Acquisition System And Compression Algorithm

Spectroscopy is a reflection of the nature of substance,and it has a high research value in target detection and recognition compared to RGB images.It has been the focus of research in recent years.Traditional spectral equipment are very bulky and can only obtain images by scanning.Only static information is feasible to be imaged,regardless of the facts that it takes a long time to do so.How to make a miniaturized,mobile spectroscopic equipment has become a bottleneck restricting the wide application of optical spectrum.Because each spectral image contains up to tens or even hundreds of channels,compared to ordinary color video,the size is enlarged dozens of times,which brings many problems for storage,transmission and processing.Especially on mobile platforms,it requires wireless transmission of spectral data.Current wireless transmission bandwidth cannot meet the requirements of real-time transmission.It is,therefore,very meaningful to study spectral compression.By far the real-time video compression methods are still aimed at three-channel RGB video.For spectral images,there is not only no specific compression method,nor a complete compression standard.The main contribution of the paper is to develop a mobile spectrum video acquisition system and discussed the application of spectral compression by ordinary video compression methods.According to the characteristics that there is a big difference between the correlation of channels and space in spectral images and those in common videos,this paper proposed a spectral compression method based on deep learning and compared it with RGB video compression method.Because of the large gap between spectral applications and traditional video applications,traditional video compression evaluation criteria does not fully represent the compression effects of the spectrum.Therefore,we raised our own set of criteria to evaluate spectral compression quality,and used compressed spectral data for applications.This paper is structured as follow:First,it introduced the origin of the problem,the characteristics of spectral imaging,the current spectral imaging methods and equipment,and the current methods for spectral compression at home and abroad.Aiming at the problem that the current spectral devices are generally bulky and unable to be moved while imaging,we developed a miniaturized spectrum acquisition system for mobile wireless transmission.Afterwards,in order to solve the difficulty of spectral video transmission and storage,a method for spectral compression was studied.Our first step was to apply RGB video compression methods to spectral video and discuss the performance of the RGB video compression algorithm in spectral compression,pointing out the shortcomings of traditional compression algorithm.Secondly,we used a deep learning method to design a multi-layered symmetric neural network by Keras.In particular,lossless compression whose inputs were multi-channel spectral images was performed in the beginning based on deep leaning.After down-sampling,features of the spectral imaging were extracted by three-dimensional convolution so the dimension could be reduced.The reduced-dimensional data was entropy-encoded and output as a binary code stream to complete the spectral compression encoding process.The decoding process is the reverse of the coding process.Through the network training above,a completed deep learning codec was obtained and compared with traditional spectral video compression algorithm.Finally,we discussed some major applications of spectral imaging.And based on the characteristics of spectral imaging,we argued that traditional PSNR and SSIM cannot fully reflect the compression effect of spectral imaging.Therefore,in this paper we proposed Mean Error of Characteristic Peak,MECP,as a criteria of evaluation and used it as an additional parameter to evaluate spectral quality.Then we utilized the compressed spectral data to complete an experiment of true-and-false face tracking recognition and compared it with the conventional ordinary video tracking algorithm.