Research On Image Processing And Face Detection In The DCT Domain

With the high-speed development of multimedia technology, its applicationis more and more extensive.Because the data amount of the multimedia data isenormous, all put forward the harsh demand to the capacity of the memory andbandwidth of transmission of the network communication.So multimedia datacompression become a key technology in the fields of multimedia, networking& communication and computer applications . Vast amounts of multimediadata like images ,video or audio data make it necessary to store and transfermultimedia data in its compressed format. However, multimedia applicationsrequire technologies permitting us to process multimedia data as flexible as wecan. For example to image, we might hope the operations such as freedomscale translation , rotate and shape vary of geometry or smoothing filtering,edge Extraction; To the video data , can carry on videos and edit in the bluescreen .To process compressed image, traditional method is based on spatialdomain, which means the image is first decoded to the spatial domain , thenprocess is done pixel by pixel ,and finally the images are encoded back to thecompressed format. But the method involves mass computation, because itincludes a full procedure of decoding and encoding, in which the DCT, IDCTcost most time. So we are thinking if we can process image in compresseddomain and that it may overcome that disadvantage. So, compressed domainprocessing of multimedia data has become an interesting research field. Thisthesis is done under the background as above.Chapter one of the thesis introduced the characteristic of the digitalimage , the redundancy of the digital image at first, then introduced the basicconcepts and development of compressed domain processing of image dataand compared it with the traditional processing model, and concluded theadvantages of compressed domain processing.Chapter two introduced variouskinds of technologies on image compressions and codes, and present areview of the data procession techniques in various compressed domains inmany literatures. At the end of the chapter ,the most widely used imagecompression algorithm—JPEG (Joint Photographic Experts Group) algorithmwas explained in detail then. Because DCT(Discreat Consine Transform) hasbeen widely used coding-standards such as JPEG, MPEG and H.261/263,it isone of the most extensive multimedia compresses technologies at present.Chapter three of the thesis firstly studied the definition that DCT and variouskinds of nature, then derive pixel add corresponding operations of the quantityadd, the quantity multiply ,the pixel add, the pixel multiply in the DCTdomain. Based on inner-block algebra ,we presentedrearrangement/resampling that can manipulate image location andtranslation.We derived various kinds of geometric transform by means ofmoving pixels within blocks in the DCT domain, and programmed realizationmirroring about horizontal axis, mirroring about vertical axis and rotating180°. At the last, we derived mixture, linear flitering, misplace ,manipulationof 8×8 clock on any point on image in the DCT domain . Because it is akind of linear operation , will not have any damage to the quality of the imagein DCT image operating data directly. Because of preventing IDCT, dataprocessing efficiency had to improve greatly. And some of algorithms can beapplied to image processing , video processing, video editing. As it is a branch of the pattern-recognition in image processing , thechapter four firstly introduced the concept of face-detection and four kinds ofmain methods of face detection ,such as knowledge-based methods, featureinvariant approaches, template matching methods and appearance-basedmethods, then put forward the difficulties of face-detection in the DCTdomain. The chapter five of the thesis presents a multimodal Gaussian modelbased approach for face detection in compressed domain, which combinesneural network classifier and skin tone verification module .The approachutilizes multimodal Gaussian model to approximate the face patternsdistribution in the feature space.The face samples are grouped into severalclusters by modified k-means algorithm. Meanwhile, the clusters of non-face...