Study On Face Recognition Based Deep Subspace Fusion Network

Image recognition is the integrated applications of computer vision and pattern recognition,which is to point to fully understand the image by extracting feature.Face recognition as an important branch of image recognition,gradually become the most popular in the field of application in the authentication and monitor the security.Deep learning model achieved a great success in the field of computer vision.Based on deep learning thought and deep subspace model which is a feature extraction model and combined with the related research results at home and abroad,this article put forward face recognition models with excellent feature extraction ability.First of all,considering that the local contact internal in image and actual environment variation,this article put forward deep subspace model under the Gabor feature descriptor modulation.The model based on a new type deep learning framework uses Gabor filter to process images,changes number of layer and the form of convolution kernels,combines global features and local features,draws a multi-layer network from the deep feature extraction built,and gets the deep abstract feature modulated by the Gabor.Experiments showed that the model combining the Gabor filter and the deep subspace have good robustness on the illumination,expression,and posture,and get better performance of feature extract.Secondly,in view of that the subspace map global feature extraction,which is easily affected by the training images,this article put forward the multi-level deep network fusion feature extraction model.Based on deep subspace model,we use "convolution-pooling" network structure,and get multi-scale abstract characteristic of samples.At the same time,the model put global characteristics as main features,supplemented by local features,and use' from coarse to thin' two-step discriminant classification strategy.Experiments show that the integration of local features and global features make the model can obtain better recognition rate,and have good robustness on the illumination,expression,posture,etc.Finally,to solve the problem of not acquiring robust image representation with the small sample set due to the lack of distribution information,this paper based deep subspace feature extraction,.using intra-personal variations,make further efforts to extract low dimension and more discriminating feature by Neighborhood Repulsed Metric Learning.Last,this paper use Collaborative Representation classification(CRC)to classify.Experiments show that study of deep model joined metric learning can effectively improve the performance and robustness of model.