Design And Implementation Of Intelligent Monitoring System Based On Improved YOLOv7 And GAN Algorithm

In the petroleum,chemical,metallurgical and other process industries,their production is mostly carried out under high temperature and pressure,flammable and explosive,toxic and hazardous high-risk environments,leading to frequent safety production accidents.Therefore,enterprises use video inspection in key parts of the plant to reduce the safety risks caused by human factors,which can effectively improve the efficiency of safe production.However,according to research,there are still many enterprises due to financial constraints,the site ordinary cameras can not be upgraded to high-definition cameras.In this paper,we design a software system for this problem,and use the improved GAN algorithm to improve the resolution of surveillance video without upgrading the camera,and then use the improved YOLOv7 algorithm to identify the target of the video processed by the GAN algorithm,and realize the software method to improve the resolution of the camera and detect the target of it,and achieve better results.The main research work of this paper is as follows:(1)MSFSA-GAN network is proposed to improve the visual effect of GAN network: multi-scale fusion is used to increase the information utilization,the low-level and high-level feature information is repeatedly acquired by parallel convolution and parallel residual recursive units,the self-attention mechanism is added to enhance the feature learning ability and the dense block extraction features are introduced to improve the loss function to enhance the visual effect.The MSDF-GAN network is proposed: the use of RRBB blocks in the generator to extract high-level features,four loss functions to reconstruct the errors.The models of both MAFSA-GAN network and MSDF-GAN network converge when compared on public dataset,and the reconstruction effect of MSDF-GAN network is better by visual effect and three metrics comparison of PSNR,SSIM and LPIPS.(2)The YOLOv7 model is combined with CAM and CBAM attention mechanisms respectively to improve small target detection,and the above two attention mechanisms are improved to L-CAM and CLAM in order to reduce the computation,and then fused with YOLOv7 model.The ablation experiments of the four improved models were conducted on the public safety cap dataset,and the YOLOv7-CLAM model was the optimal model with speed and accuracy of 46 s and 74.36%,respectively,which is 5.83%higher accuracy and 86 s higher speed relative to YOLOv7,and the number of parameters only increased by 1.3M,indicating the effectiveness of the improved model.(3)Combining MSDF-GAN with YOLOv7-CLAM model to develop intelligent monitoring system to obtain MGYC model,which is trained after Make Sense labeled homemade dataset,and the accuracy of training and test sets are finally stabilized at 91%and 88%,with a speed improvement of 9.92s～26.15 s and an average accuracy improvement of 4.82%～17.98% compared with other classical models.It shows that the fused model is effective.In addition,the SR intelligent monitoring system was developed using Py Qt5 interface development library and My SQL database to achieve real-time recognition of people,cell phones,flames,smoking and whether they are wearing helmets,and voice alarm for abnormal targets.