Research And Application Of Object Detection Guided By Attention Model

The technology of locating objects in images or videos is called object detection technology,and it has important application value in daily life and industrial production.Generally,the target detection network based on deep learning has a more accurate recognition rate and faster recognition speed than traditional image processing methods.In recent years,convolutional neural networks in the field of deep learning have developed rapidly.Nowadays,detection technology based on deep learning has been widely used in many fields.For example,detecting pedestrians on the road,detecting the wearing condition of motorcyclists' helmets,and so on.In order to meet the needs of various applications,target detection algorithms are constantly improving.In this paper,the SSD(Single Shot Multi Box Detector)network is used as the basic network for research,and the guiding attention model is added in.Using the improved network to handle motorcycle helmet wearing detection problem.Discuss the realization principle of the SSD-Net and some current mainstream attention models from the shallower to the deeper.When building the specific implementation of the network,the module of one-step detection network SSD-Net and visual attention mechanism is properly integrated first,then through the lightweight processing of the network to improve the network.The research content of this paper is as follows:(1)Firstly,SSD-Net network is used to detect and optimize the helmet wearing condition of motorcycle drivers.Use the designed network to training the motorcycle driving photos collected on the road through the cosine decay learning rate.The network model is obtained by adding the RFB(Receptive Field Block)module to the SSD-Net,which uses VGG16-Net as the basic network to extract features.The design is carried out on Ubuntu 16.04 system through the Pytorch framework.The performance index MAP value of the experimental results shows that the RFB-NET model is better for the task of detecting whether a motorcycle driver is wearing a helmet,which has a higher accuracy rate.(2)Secondly,In order to detect whether motorcyclists wear helmets,a module similar to visual mechanism is introduced into the mainstream single-stage detection network SSD-Net.The weight of the network feature map was re-selected in channel and space,and the RFB module similar to human visual eccentricity mechanism was added.In addition,Mosaic method was used for data enhancement,and cosine attenuation learning rate was used for optimization to complete the detection of whether motorcyclists wear helmets.The test results show that the MAP of the improved network on the motorcycle helmet wearing detection task is about 4%-5%higher than the original SSD-Net,which shows that the new network has better application effects.(3)Finally,the lightweight backbone network Mobile-Net is selected to make the detection network lightweight,and the lightweight detection network is used for the standardized detection of motorcycle riders' helmet wearing.By replacing the main network VGG-Net of SSD-Net with Mobile-Net to realize the MobileAttentional-SSD-Net.Under the premise of ensuring detection accuracy,this network reduces the number of network parameters,saves storage space,and accelerates the detection speed of the network,which shows obvious advantages in real applications.