| Human post estimation and recognition has always been a challenging subject in computer vision.In recent years,the rapid development of deep learning technology has greatly promoted the research in this field.Most of the algorithms focus on improving the recognition accuracy,which makes the computational complexity too high.Most of these algorithms use multi-stage network model.Although the precision of the network structure is improved,the parameters of the trained model are very large.Both operation and deployment are greatly dependent on the investment of hardware cost.The complexity of these algorithms mainly comes from the design complexity of the model itself and the deep learning convolutional network parameter calculation.Based on deep learning technology,this paper aims to optimize and refine network structure,reduce network parameters,and build a lightweight and accurate human post estimation algorithm under the condition of keeping algorithm accuracy as far as possible.Firstly,deep separation convolution is used to replace the ordinary convolution computation,thus reducing the number of parameters of the whole network.Design end-toend network at the same time to the human body target were analyzed,and further reduce the size of deep learning network,in terms of feature extraction,in order to reduce the depth of the separation of convolution with the feature of ability is insufficient,the effect of using efficient hourglass model the feature pyramid network structure,finally,the model of quantity and precision are analyzed in detail,validate this algorithm to reduce the network parameters and maintain the effectiveness of the two aspects of accuracy.The following is the main content of this paper:1)From the perspective of image and video sequence,single target and complex target,this paper expounds the current research status of human post estimation and recognition,and analyzes the purpose and significance of the research based on the existing algorithms.2)Analyze the convolutional neural network technology,compare the difference between the traditional image processing algorithm and the image processing algorithm based on the convolutional neural network,and elaborate the technical means of the lightweight network and the high-efficiency feature extraction technology.3)Analyze the coding method of human post estimation,compare the performance difference between top-down and bottom-up algorithm,and point out the research emphasis and improvement methods of the two type algorithms.4)In the network for extracting key points of human post,pyramid residual network based on hourglass structure is applied.Multi-scale features are fully considered in the design of feature extraction network,so as to maintain sufficient detection features when network parameters are reduced and to improve the impact of insufficient features brought by separation and convolution;The heat map was used to detect the key points and part of the affinity field was used to solve the key points planning problem in multi-person post estimation.5)Experiments verify that the combination of Depth-wise separable convolution and feature pyramid network with hourglass model can not only reduce network parameters,but also maintain high identification accuracy. |
