Small Sample Face Recognition Based On Collaborative Representation And Siamese Network

With the the development of modern information society,the problems of information security and accurate and efficient identity authentication have been pushed to a very important position.The traditional identification technology relying on password or physical media can't meet the current needs.Face recognition,as one category of biometric recognition,is widely used in video surveillance,financial products,security,military,and other fields because of its uniqueness and difficulty in reproduction.In real life,face recognition would be influenced by various factors such as the complex background,lighting,face expression change,obstruction,age change.Therefore,the algorithms are required to have certain robustness to external factors.In addition to the problems above,there are some cases where it is not possible to acquire enough training samples for a single category in face recognition applications.It is difficult to learn effective features with insufficient face samples.Sparse representation is widely used in image processing and pattern recognition.The collaborative representation has an algebraic solution,which greatly reduces the amount of computation and the recognition performance is quite similar.Therefore,the face recognition algorithm based on collaborative representation is studied in this paper.Combining the local structure with multi-scale,it still has excellent identify performance when the number of training samples is small.Compared with the traditional method,deep learning does not need to extract features manually,which reduces the loss of features to a certain extent.The convolutional neural network has a particularly outstanding ability in image processing.Therefore,this paper deeply studies the theoretical knowledge of convolutional neural networks.The existing convolutional neural network is improved and a network structure suitable for face recognition is proposed.The work of this paper is mainly described as follows:(1)In the case of a small sample,the recognition performance of collaboration expressions can be severely affected.The multi-scale block cooperative representation algorithm can effectively integrate the classification results under different scales,but the calculation of sub-blocks in the classification framework is independent to each other,which results in ignoring the structural relationship between the blocks.The local structure method divides the image into multiple local regions.The overlapping blocks of each local region are distributed in the same linear subspace.This subspace can reflect the structural relationship between the blocks and improve the multi-scale block cooperative representation in small Robustness under the sample.The experiments show that the multi-scale block co-representation based on local structure can effectively improve the recognition accuracy when the number of training samples is small.(2)When enough training samples is not available for every single category,it is difficult to learn effective features using the deep learning method,which results in the unsatisfied recognition performance of the network model.Therefore,this paper proposes a facial recognition algorithm based on a convolutional neural network.The main work has the following points:? We designed and built a new Siamese network model SiameseFacel.The face image pair is used as the input of the Siamese network and the input image pair is mapped into the target space so that its L2 norm distance in the target space can represent the semantic distance of the input space.The mapping is expressed through a neural network that is supervised and learned.The experimental results show that SiameseFacel achieves high accuracy in both AR dataset and LFW dataset without using external datasets.Another new and more lightweight Siamese network model SiameseFace2 was designed and implemented.The network structure of this model is more optimal.The 1*1 convolution kernel is used to reduce the computational complexity.The underlying features are extracted and cascaded with high-level features,making the model more accurate.SiameseFace2 not only greatly reduces the network model parameters,but also does not lose face recognition accuracy.? In the currently disclosed face datasets,the number of face samples per class is relatively small.For deep learning,it is difficult to train an excellent network model without sufficient data.To solve the limitations above,this paper proposes a new training data generation algorithm,which indirectly greatly expands the number of training samples for a single category of AR and LFW and alleviates the drawbacks of fewer face images of a single category to a certain extent.Using this data generation method,the recognition accuracy of the model is significantly improved.? This paper attempts a number of different loss functions,including triplet loss function,square loss function,cosine loss function,and the contrast loss function used in this paper.Through comparison experiments,the best loss function with the framework of this paper was selected.