Research On Multi-target Detection And Recognition System In Intelligent Vision Sensor Network

With the advent of 5G and AI eras,the intelligent visual sensor network system has developed rapidly,showing great potential in many fields such as new retail,urban brain,industrial Internet of Things,etc.,but deep learning relies on massive data,and the cost of collecting labels is high;The general application scenario requires high-precision algorithms,low-delay feedback for detection results,and support for large-scale identification libraries.In this paper,the architecture model combining cloud and edge computing is proposed.Some computational forces are forwarded to the edge computing processor to complete the target detection work for massive unstructured image data,and then send the detection target to the cloud recognition processing.This approach reduces latency,improves recognition accuracy,and supports the deployment of large-scale identification libraries.The data required for the whole detection and identification algorithm training is obtained through a synthesis scheme,which reduces the data collection cost.The main methods are as follows:The data synthesis scheme includes: image segmentation network M-Unet algorithm,image synthesis strategy,CycleGan enhanced network algorithm,M-Unet adds multi-path Dilated Convolution module(MPDC)compared to traditional U-Net network,Global Pooling Network(GCN),Soft Dice loss function,improve the segmentation performance of different size targets;synthesis strategy sets the synthesis rules through four parameters: angle,occlusion,scale,and background;the synthesized image istransformedbyCycleGanfor style conversion,and finally obtains the data combination of the real scene,which greatly reduces the cost of collecting and labeling,and is used for subsequent detection and identification of services,which has great practical application value.In the multi-target detection method of visual sensor terminal,the implementation process of the two-stage target detection algorithm FPN is introduced,and the M-FPN based on the positive and negative sample optimization strategy is proposed.The specific method is to improve the IoU screening threshold(ie,the positive sample quality)while the IoU screening threshold(ie,the positive sample quality)is disturbed,thereby increasing the number of highquality samples and verifying the average detection accuracy on the MS-COCO data set.The mmAP increase of 1.2% demonstrates the effectiveness of the optimization strategy.Then it further proves the feasibility of the synthetic data in the target detection and identification business.The average indicators of mAP50,cAcc,mCIoU,etc.in the data set reach 96.63%,56.92%,93.32%,and the relative direct synthesis schemes increase by 15.9%,47.43%,23.52% respectively the lifting effect is remarkable.The ResNet_FPM fine-grained classification algorithm is based on ResNet-50 as the basic network.The FPM(Feature Part Module)component-based feature module is added to cut the feature vector into multiple partitions.Each sub-feature tensor will be globally averaged.Average Pooling),Global Max Pooling,Fully-Connected Pooling operation to form a vector of D dimensions,and then splicing each module to obtain the final feature vector for classification,and finally increase the recognition accuracy 4.1%;Finally,the feature pyramid method is proposed,and the recognition accuracy is improved by 4.2%,which significantly improves the accuracy of fine-grained recognition to 99.8%.