Research On Face Detection Method In Complex Environment Based On Deep Learning

With the rapid development of various face applications such as face recognition,annotation,and retrieval,face detection has become more and more important in the field of image analysis,and has become one of the most studied topics in the past decades.However,due to the effects of illumination changes,different posture expressions,and various scales of human faces,face detection still faces many challenges in practical applications.In particular,images captured under low light conditions have low visibility and signal-to-noise ratio,which seriously affect the results of face detection.As a result,the performance of face detection is difficult to satisfy the requirements of practical applications.This thesis studies the relevant theories and methods of face detection based on deep learning in complex environments.It mainly focuses on low-light image enhancement methods and multi-scale face detection methods based on deep learning.It has achieved certain results.The main work of this article is as follows:1.Using an improved low-light image enhancement algorithm based on dark channel priors achieve efficient image preprocessing.For the problems of the low illumination image enhancement algorithm based on dark channel priors,which is prone to halo phenomenon and high computational complexity,an improved method is proposed.By combining the image inversion with dehaze algorithm using dark channel prior,the global atmospheric light and transmission maps are directly estimated on the original low-illumination image.And the estimated rough transmission maps is refined by the guided filtering algorithm.The generation of halo in the image is suppressed,and the running speed of the algorithm is improved.2.The convolutional neural network based on the region Softmax loss function is designed for image classification.Aiming at the problem that the loss function of the traditional convolutional neural network is only related to the global information of the image but lacks the ability to fit the image subregion,the region Softmax loss function is proposed.According to the mapping relationship between the convolutional feature map and the image sub-regions,the characteristics of each channel in the convolutional feature tensorare mapped as local regions of the image,thereby enhancing the understanding of the sub-regions by minimizing the loss of each channel feature.Therefore,visual representation ability of the network is improved.3.A multi-scale face detection method based on convolutional neural network is designed to implement face detection.Since there are only two categories of face and non-face in face detection,the category labels of the objrction and the background in the objection detection are similar.Therefore,the region proposal network in the Faster R-CNN could be use for face detection.And for the problem that the region proposal network easily ignores small-size human faces,face detection is performed on feature map from different convolutional layers through three sub-detectors with different anchor-point scales.And the combination of the region Softmax loss function with sub-detector to improve the result of large scale face.The experiment show that the proposed face detection method has achieved advanced results on multiple face detection datasets.In the end,the method proposed in this thesis is designed and implemented in a complex environment including low-light image enhancement and face detection.The face detection system can achieve accurate positioning of human face in low illumination environment.