Research On Anomaly Target Detection Algorithm Of Urban Management Image Based On Deep Learning

In the past 10 years,with the rapid development of computer vision,deep learning has also been rising and developing,and has made rapid progress in the areas of community security,mall monitoring,city management,and detection of prohibited items in important transportation hubs.This paper mainly aims at four kinds of targets,namely road garbage,flow operation,outbound operation and illegal parking,to achieve the target detection of small scene images shot by grid personnel’s mobile phone and large scene images shot by fixed camera.The following aspects are mainly studied:(1)This paper collates and annotates four types of image data sets with abnormal targets in urban management captured by grid personnel using mobile phones,and proposes an abnormal target detection algorithm for small scene images based on deep learning city management mobile phones.The algorithm uses the improved YOLOv3 network to locate and classify abnormal objects in the image.The residual block in the network architecture is changed to a convolutional cascading residual block.The activation function is changed to h-swish.The four parameters of the border regression function are added to Gaussian.The distribution constitutes the frame regression loss function,and the four parameters are the coordinate value and width-height value of the center point of the frame.In addition,difficult sample mining is used to iteratively update the model to improve the detection accuracy of the model.(2)Aiming at the surveillance video image data set,this paper proposes an image abnormal target detection algorithm based on deep learning city management video large scene.Due to the difficulty of collecting images with abnormal targets in the surveillance video,this article uses the abnormal images captured by the grid personnel’s mobile phone to enhance the video images to construct a training data set for video abnormal target detection.The video image data enhancement methods are pixel amplification on the image taken by the mobile phone,pasting the image taken by the mobile phone into the video image,and pasting the four types of target images from the mobile phone image and pasting the target onto the video image.A multiscale feature pyramid output network architecture with cross-layer connection is designed for anomaly target detection in city management surveillance video.The trained model was tested separately on the data set constructed by ourselves and the public data set.The experimental results show that the first algorithm proposed in this paper improves the detection accuracy of the images taken by the four types of city management mobile phones,and the model has high robustness.The accuracy of target detection has been greatly improved.