Research On Identification Of Illegal Business Behaviors Based On Mask R-CNN

In recent years,computer vision technology has made great progress.In addition to medical detection,video surveillance and other fields,it also has broad application prospects in urban management,such as the use of instance segmentation to identify urban illegal business behaviors,making urban management more efficient intelligent.However,there are the following problems:(1)The detection accuracy of instance segmentation models such as Mask R-CNN is low,especially for objects with uncertain forms such as illegal business behaviors,the detection accuracy is low,the mask segmentation accuracy is low,and the object contour boundary segmentation is not clear enough;(2)There are interference factors such as fog and haze in the real environment,which have a certain impact on the detection ability of the model;(3)There is no public data set of illegal business behaviors,which is inconvenient for the training of neural network models;(4)There is currently no effective system for identifying illegal business operations to promote more efficient city management.Therefore,based on the above background,this thesis mainly studies how to improve the accuracy of model object detection and mask segmentation by improving feature reuse and feature fusion,in order to build a system that can effectively identify illegal business operations in cities.In this thesis,Mask R-CNN is used as the basic architecture,and a network architecture with instance segmentation function DU-AFNet(Down to Up-Weight Fusion neural Network)is constructed.The main contributions are as follows:(1)According to the idea that different scale features have different effects on the formation of the final feature,Mask R-CNN is improved,a bottom-up feature fusion path is constructed to achieve multi-scale feature fusion,and combined with convolution operations,adaptive multi-scale feature fusion,obtain the fused feature map by taking the maximum value,and then use the feature map to perform subsequent object classification tasks and mask segmentation tasks;in the mask generation branch,add an optimized fully connected condition random field to improve model mask segmentation Ability;(2)In the DU-AFNet network architecture integrate the end-to-end defogging module AOD-Net to enhance image clarity,and improve the detection accuracy of the model in severe weather;(3)Built a city illegal business behavior data set: vendor data set;(4)Build a system that can automatically identify illegal business behavior.In this thesis,a comprehensive experimental analysis of the improved model is performed.The experimental results show that compared with Mask R-CNN,only the bottom-up model with adaptive multi-scale feature fusion paths is added,and the AP value of the detection object frame is improved by 1.3% in the COCO data set,an increase of 3.4% in the vendor data set;a model that only adds optimized fully connected conditional random fields,the AP value of its mask segmentation is increased by 1.1% in the COCO data set,and 4.3% in the vendor data set;DU-WFNet integrates the above two improved modules,the AP index on the COCO data set has increased by 1.2%,and the AP on the vendor data set has increased by 2.7%;after adding DU-AFNet to the AOD-Net defogging module,the AP value of the model has increased by 0.9 %,and training time only increased the time cost of 0.7%.The experimental results show that the illegal business behavior recognition system based on the above neural network can effectively simplify the city management.