Research On Similarity Judgment And Detection On Face Images Based On Lightweight Network

Face recognition technology has a long history and is now gradually maturing,and has been widely applied in various fields.Due to the development of face recognition technology,face similarity determination and facial image detection technology have also received attention in many fields,such as criminal investigation,criminal target matching,missing population location and some security technologies,etc.The facial image of the human body is a numerical display of biological features.We can obtain the biological features of the image through the concept of mathematical function,and use this function to calculate the image features to achieve the effects of recognition,detection,segmentation,etc.However,the mathematics of facial images features are also disturbed by a series of factors such as shooting angle,age,light,background and so on.In determining the similarity of facial images,the traditional algorithm mainly obtains the geometric structure data of the face contour based on the theoretical knowledge of topology,and then analyzes the topological information of the geometric data of the facial facial organs to obtain between the image pairs similarity value;in face image detection,most of the traditional target detection algorithms are based on Haar for feature extraction,and then combined with Adaboost to achieve the classification effect.However,in the face similarity judgment,the accuracy of these traditional algorithms will be seriously reduced due to the interference of factors such as too complex background,multiple angles of the face,and large age span of the face.Also in face image detection,the limitations of traditional algorithms will also be affected by practical problems such as image background complexity,shooting angle,etc.,resulting in problems such as slow detection speed,low accuracy,and low detection efficiency.In the field of computer vision,the most popular image processing algorithm is the convolutional neural network.Convolutional neural networks also give classic algorithm twin network models for determining the similarity of facial images;some classic algorithms for facial image detection.Compared with traditional algorithms,these new algorithms based on convolutional neural networks have great breakthrough effects,but because of the interference of the objective factors of the models in convolutional neural networks,these new algorithms also have many limitations.Here we make the following improvements based on the convolutional neural network:(1)Research on Double-line Lightweight Twin Network.Here we mainly put forward an argument on the basic structure of the twin network: combining traditional image processing algorithms with a simple convolutional neural network to achieve the lightweight effect of the twin network.The main purpose of the image processing algorithm here is to extract the edge features of the image.We first conduct experiments on Canny,LBP,HOG,Robinson and other operators.The experiment obtains the superiority of the Robinson operator,and at the same time,we do some parameter improvement on Robinson to obtain the NR operator.Compared with the edge feature extraction effect in the existing literature,the NR operator intuitively draws the advantages of the NR operator in edge feature extraction.Here continue to combine the NR operator with a simple convolutional neural network to design a lightweight twin network SiaR Net model.The experiment proves that the SiaR Net model has advantages.(2)Research on the three-wire lightweight twin network.The superiority of SiaR Net unilaterally demonstrates that there is certain superiority in the combination of traditional image processing algorithms and simple convolutional neural networks to achieve the lightweight effect of twin networks.Here we apply this argument to the three-wire twin network.At the same time,based on the network structure of the three-wire twin network,the relative discrete function of the image is proposed,and finally the lightweight three-wire twin network NTn model is designed.The relative discrete function of the image is mainly to use the structural basis of the three-wire twin network to replace the previous position of the Euclidean distance function in the twin network.This paper makes a comprehensive comparison experiment between the NTn model and the Squeeze Network from the two data sets CACD2000 and VGG_FACE2.Compared with the Squeeze Network in terms of model accuracy,the NTn model has been improved by 11%;compared with the Squeeze Network in terms of the model size,the NTn model gets a reduction of 16.4MB.(3)Efficient research of twin network in target detection network.In view of the low efficiency of traditional target detection in special environments such as judicial reconnaissance,criminal target detection,missing population detection,etc.,this paper proposes an efficient target detection algorithm,MT-Siam which is to be able to quickly and accurately obtain the specific position of the target person in the image.In the face of the classic target detection networks YOLO,MTCNN,and SSD,here we first analyze the methods of generating candidate frames and the extraction of facial features and specific processing methods of these detection networks during the detection process,and then compare the detection effect of them under the FDDB data set.From the detection effect,it is concluded that the accuracy rate of MTCNN is 86.7% better than YOLO and SSD,and the model size is 2.1MB better than SSD and YOLO.Considering the accuracy and model size,the structure of MTCNN model is more suitable as the basic testing skeleton of efficient target detection.In order to realize the efficient detection model MT-Siam in combination with the MTCNN model,here we train the dual-line lightweight twin network Siam under the CACD2000 and VGG_FACE2 data sets to improve the generalization ability in similarity determination,and then train the trained Siam model which as a similarity judgment factor,the MT-Siam model is designed in combination with MTCNN.It is experimentally verified that MT-Siam is superior to MTCNN in terms of efficiency.The model is reduced by 15%,and the detection speed is increased by 84%.And get a usability argument for the above argument.