Research On Facial Expression Recognition Algorithm Based On Deep Learning

Pedestrian detection is an important branch of object detection and computer vision.Relevant research results have been widely used in intelligent monitoring,intelligent transportation,vehicle assisted driving and other specific scenarios.Due to the uneven illumination,occlusion,noise and other factors in the complex scene,the current pedestrian detection algorithm still faces the challenges of low detection rate and insensitivity to small targets in the application of the actual scene.Based on convolutional neural network,this paper constructs a faster and more accurate pedestrian detection algorithm,which provides a new method for pedestrian detection in complex scenes.First of all,this paper introduces the research status of pedestrian detection algorithms at home and abroad,and also analyzes several major difficulties faced by the current pedestrian detection task.At the same time,the advantages and disadvantages of the method based on artificial feature representation and classifier and the method based on depth learning are introduced in detail.The common data sets and evaluation criteria of pedestrian detection algorithm are also introduced.Secondly,on the basis of fast RCNN algorithm based on deep learning,this paper improves it according to the characteristics of pedestrian data set.For multi-scale pedestrian target,three sliding windows of different sizes are used to generate candidate regions to obtain better receptive field.In order to improve the ability of multi-scale target detection,the method of feature stitching is used to fuse the features of shallow details and deep high semantics in convolution network.Finally,the algorithm is tested on VOC 2007 dataset in the open source deep learning framework.The experimental results show that the improved measures based on fast RCNN can effectively improve the detection ability of pedestrian detection system in relevant data sets.Finally,the fast RCNN algorithm is improved according to the characteristics of Caltech data set of vehicle video in traffic scene.Because the task of pedestrian detection is a two classification problem in nature,the number of output neurons inthe full connection layer is only 2048,which reduces the system over fitting.In order to better adapt to the characteristics of length width ratio of pedestrians in Caltech data set,the characteristic length width after regional pooling is respectively 8×4.In order to improve the detection ability of the network,this paper adds a method of region pooling around the candidate window.The experimental results on Caltech data set show that the improved method can achieve higher detection accuracy.