A Lightweight Algorithm Based On Multi-Scale Awareness For Face Detection And Crowd Counting

With the gradual maturity of 5G technology,the era of the Internet of Things with 5G as the core is coming.Cloud computing and edge computing will jointly accelerate the development and integration of the Internet of Things and artificial intelligence.The construction of information society and smart city requires the high precision artificial intelligence algorithm in discovering social security risks and fighting illegal crimes.More and more security surveillance equipment produce massive amounts of image and video data.How to process these data efficiently,accurately and intelligently through artificial intelligence technology is still a challenging research direction.This paper proposes a lightweight convolutional neural network algorithm in edge computing environments for face detection and crowd counting tasks in dense crowd scenarios.The experiments have proved that the algorithm has excellent detection accuracy while maintaining high efficiency and low latency.Limited by the computing power of edge terminals devices,it is difficult to deploy face detection algorithms with complex structures.In order to reduce consumption resource and ensure detection accuracy in complex scenes such as multi-scale changes of face and crowd characteristics,occlusion,blur,and illumination,this paper proposes a scale-aware lightweight face detection algorithm(Scale-aware Dual Path Network,SDPN).The improved face residual neural network called Face-ResNet is used as the feature extraction backbone,and the DPE(Dual Path Shallow Feature Extractor)is proposed.The parallel branch understands the multi-scale image information,and then the deep and shallow feature fusion module combines the low-level image information with high-level semantic features.Then the multi-scale awareness training strategy supervise the multi-branch learning discriminating features.Experimental results show that the proposed algorithm can effectively extract diversified features,while maintaining the efficiency of inference and low latency,it effectively improves the accuracy and robustness of the algorithm.Furthermore,this paper proposes the SA-SDPN(Separable Attention Scale-aware Dual Path Network)algorithm for the task of crowd counting.The design ideas and core concepts of the face detection network in dense scenes are combined with the crowd counting task,and the face residual network is used as the backbone feature network in the crowd counting task.Based on this,the introduction of dilated convolution obtain a wider range of sparse and dense features while retaining rich spatial feature information.The introduction of a separable attention mechanism layer with spatial dimensions and channel layers explicitly optimizes the expression of the local and global correlation characteristics of the network.The crowd counting network proposed in this chapter is improved based on the MCNN algorithm framework in terms of probability density map generation,data preprocessing and network design.In the experimental stage of this chapter,a sufficient sub-module comparison validation has been carried out to prove the effectiveness of the improvement,and a comparison test with other algorithm has been carried out.In the end,the validation in The Shanghai Tech dataset proves the superiority of the proposed algorithm over other algorithms.