Research Of Face Detection Technology Based On Deep Learning

For the face detection problem with multiple attitudes,illumination effects,occlusion and other factors,deep learning detection algorithms have become the mainstream,and most algorithms are accompanied by deeper and more complex networks at the cost of improved performance.Because the network classification ability is too small,the large network classification ability will increase the computational cost and the parameters will be easily overfitted.This paper studies the performance of different CNN algorithms on face detection tasks.Firstly,the advantages and disadvantages of existing face detection algorithms are analyzed.Taking into account the requirements of accuracy and effectiveness of face detection tasks,the AlexNet and VGG models and DeepID models are compared and analyzed,and the DeepID model and AlexNet model are selected as research objects.In order to explore the performance differences between different network depths and the number of convolution kernels in face detection,this paper improves the network structure based on the DeepID model,and designs convolutional neural networks under different configurations.Through the comparative analysis of network performance under different configurations,the factors affecting the network performance are derived.At the same time,the performance of different networks under the influence of front,side,and non-uniform illumination is analyzed.Aiming at the problem of too many sliding windows for traditional CNN detection algorithms,this paper improves the trained convolution model by using full convolution.It improves the disadvantage that the traditional convolution model can only accept fixed-size input and avoids the redundant calculation of the traditional sliding window method.,improve the detection speed.Aiming at the problem that the parameters of AlexNet network in the task of this article are easy to overfit,this paper proposes an improved method including the optimization of network structure and the improvement of training strategy.The network structure first compresses the full connection layer to reduce the calculation parameters,and then uses the cascade.The small-scale convolution kernel replaces large-scale convolution kernels,while multi-size convolution kernels are introduced to carry out convolutional feature fusion of different scales and use 1×1 convolution to check parametric dimension reduction.The training strategy proposes a combination of data enhancement and step-by-step training.Experiments show that after the improvement,the network detection performance is improved and the false detection rate is reduced.For the problem of slow detection and inaccurate face location in single-level convolutional neural networks,this paper studies and designs a cascaded convolutional neural network as a face detection algorithm and analyzes its detection performance.The experimental results show that the cascaded network speeds up the detection speed while reducing the false detection rate and more accurate face positioning.Finally,the robustness,accuracy and timeliness of the three different face detectors designed in this paper are experimentally analyzed.The results show that the cascaded detector performs better than the single-level detector.The detection rate on the FDDB data set is 80.9%,and the detection rate on the LFW data set is 96.9%.When the image resolution is less than 300×500,The network algorithm can achieve the effect of real-time detection.