Research On Face Detection Based On Lightweight Convolutional Neural Networks

With the rapid development of mobile internet and artificial intelligence,various intelligent terminal products have entered the public’s vision.As an important component of facial detection,it is widely used in various fields such as monitoring systems,intelligent transportation,and live detection.The face detection task in practical applications needs to be based on mobile terminal platforms or embedded platforms,with limited computing resources.However,traditional face detection models have a large computational load,complex structure,and slow detection speed,which cannot meet the needs of porting to low-power mobile terminal platforms or embedded platforms.Therefore,It is crucial to develop high-precision and lightweight face detection algorithms that can run in real-time on low-power platforms such as mobile terminals or embedded systems.This article is dedicated to the research of high-precision,lightweight,and real-time facial detection algorithms.On the basis of summarizing the current popular lightweight Convolutional neural networks and object detection algorithms,an improved lightweight Convolutional neural network face detection algorithm Mobile SSD is proposed by analyzing the network framework and its advantages and disadvantages of the SSD algorithm model.It can effectively reduce parameter and computational complexity while ensuring detection accuracy,and is more suitable for resource constrained mobile terminal platforms.The specific research results are as follows:1.In response to the problems of large size,multiple parameters,and high computational complexity of the backbone feature extraction network VGG16 in the SSD algorithm model,a lightweight linear transformation feature map is used to replace the original Convolutional feature map,and VGG16 is replaced with a lightweight Mobile Net V2.Compared with the original Convolutional network,the lightweight linear transformation of Mobile SSD can effectively reduce the model size,number of parameters,and computational complexity while generating the same number of feature maps.2.To solve the problem of dead neurons in the Re LU activation function in the SSD algorithm model,use a more efficient PRe LU activation function to replace the original Re LU activation function.The PRe LU activation function used by Mobile SSD can solve the problem of dead neurons,and at the same time let the neural network adaptively select parameters,which speeds up the Convergence speed of the network.3.In view of the imbalance of training samples in the Softmax Loss loss function in the SSD algorithm model,Focal Loss loss function is used to replace the original Softmax Loss loss function.The Focal Loss loss function used by Mobile SSD can ensure that there are enough positive samples and hard to classify samples to dominate the gradient update direction of the loss value during the model training process,so as to train the model more fully.4.In response to the issue of insufficient consideration of the spatial relationship between different Convolutional layers in the SSD algorithm model,which leads to unsatisfactory detection of small targets,a multi-scale feature fusion module is used to enhance the detection of small targets.The design of multi-scale feature fusion used by Mobile SSD fully utilizes feature information from different levels to achieve the goal of fusing the advantages between features,thereby improving the performance of the model and strengthening the detection of small targets.