Font Size: a A A

The Video Face Recognition Method Based On The Deep Learning

Posted on:2014-04-06Degree:MasterType:Thesis
Country:ChinaCandidate:Q Z YouFull Text:PDF
GTID:2268330392969571Subject:Computer technology
Abstract/Summary:PDF Full Text Request
The basic design idea of the video face identification and detection methods is:after the video files and their subtitles and scripts are given, it can automatically detectand identify the characters in the video, does not require any training samples. Videoface recognition and detection method mainly consists of four parts: subtitles andscreenplay fusion part, face detection portion, the sample set automatically generatedpart and face recognition part based on deep learning. This paper introduces depthlearning algorithm into the video face recognition, there are two aspects’ importantsignificance. The one hand, the video face recognition algorithms has certainanti-jamming capability, and can guarantee the real-time, the experiments and analysisshow that the depth learning algorithm with these requirements; On the other hand,from the point of view of the characteristics of depth learning algorithm, the biggestdrawback of depth learning algorithm is that depth model requires a large number ofsamples, which largely limits the application of the depth learning algorithm. However,this designed video-based face detection module in this paper can easily generate tens ofthousands, hundreds of thousands of samples to meet the large sample set requirementsof the depth learning algorithm.The face recognition part based on the depth learning model is the core of theentire system. The significance of this part consist of two aspects: first, after the videoface detection part, although the purity of the human face in the video face collectionhas been greatly improved, but still there are some impurities, therefore the recognitionmodule must be used to further filter out the impurities in the collection of human face;second, through the frame files obtained from the video, at the same time more than oneface occur is possible, and in this case, video face detection section cannot handle thespeaker corresponding to the face, the identification module must be used to distinguishmore than one face in one frame.The face recognition part based on depth learning model mainly consists of threemodules: data preprocessing module, depth learning modules, and recognition module.Data preprocessing module mainly consist of the data integration and structure data twoparts. Depth learning module consists of two parts: RBM regulation and feedbackfine-tuning of the depth model. The adjustment process of RBM is the adjustmentprocess between the respective layers of the bottom-up, in this way to initialize theweights of the entire depth model system. The feedback fine tuning of the Depth model,firstly, the bottom-up recognition model conversion, then the top-down generationmodel conversion, and finally through the continuous adjustment between the different levels, the generated model can reconstruct the original sample which has a lower error.This essential characteristics of this sample are gotten, sp is the maximum abstractrepresentation layer of the depth model. After the treatment of deep learning model, thecharacteristics of the samples after dimensionality reduction can be gotten, and then theidentification module is used. This paper uses the artificial neural network method to dothe Identification.
Keywords/Search Tags:face detection, skin color model, deep learning, recognition model, generated model, artificial neural networks
PDF Full Text Request
Related items