Research On Face Matching Deep Learning Algorithm Applied To Classroom Scenes

Classroom attendance of college students has severely occupied the study time,the examination verification increases the workload of the invigilator and is inefficient.With the research of artificial intelligence technology and the promotion of technical application fields,problems encountered in different fields are attempted to solve by artificial intelligence methods,especially in computer vision,thanks to the powerful computational ability of the convolutional neural network and the characteristics of nonlinear autonomous learning features,it is widely used in object recognition and verification in real-world scenarios.In order to save a lot of manpower,material resources and improve work efficiency,a deep learning method can be used to solve the face matching problem of the classroom scene.With the deep learning theory,this paper designs a face image classification algorithm based on the dataset from classroom scene.The design idea of algorithm is mainly explained from three parts: face image preprocessing,development and testing of face image classification algorithm and experimental analysis of training result.Firstly,the large scene image is collected from the classroom,the face detection algorithm is used to capture the face image,and the captured image is classified into the experimental image used as experimental dataset.For the problem of the poor quality of the image sample of the real scene,this essay used the histogram equalization and the sharpening enhancement algorithm to processes the face image,filtering the influence of environmental factors and effectively highlights the face information in the image,and performed face alignment in order to make the model more focused on learning the discriminative features in the canonical region.Secondly,this paper focuses on the face image classification algorithm designation,the algorithm consists of a classification model and a classifier Softmax,in which three model structures are designed: lightweight convolutional network model,Shortcut-based network model and improved Siamese structure model.Considering that the experimental dataset is a realistic scene sample,and for increasing the learning difficulty and obtaining more discriminative facial features,this paper focuses on improving the original Softmax of the algorithm,using Softmax based on cosine margin as the objective function.Through experimental tests,the lightweight convolutional network is the most flexible.The Shortcut-based network has the most stable performance on different datasets.The improved Siamese network improves the accuracy of the original Siamese network and has a certain degree of advantages on small datasets.Softmax based on cosine margin greatly promotes the learning process of the model for low quality samples obtained from real scenes.From the analysis of experimental results,this paper recommends a combination algorithm based on Shortcut module and Softmax based on cosine margin.The algorithm is applied in the case where the real scene environment is complex and the sample collection is not good.Lastly,this paper tested many combination algorithms which consist of the different models and different Softmax.The accuracy and Loss trend of each algorithm are observed during the training process.The advantages and disadvantages of each algorithm scheme and the characteristics of the applicable dataset are analyzed.In order to verify the generalization performance of the algorithm and the discriminability of the extracted features,this paper visualizes the 128-dimensional face feature expression vector predicted by the algorithm,observes the difference between the objects,and uses the cosine similarity formula to calculate the similarity.For untrained identity samples,the algorithm can still match accurately,and the experimental results further verified the effectiveness of the algorithm.