Research On Deep Learning Method Of Smart Urban Management Scene Understanding

In recent years,along with the rapid development of the city,further request for urban management is put forward.The traditional "Digital Urban Management" has gradually evolved into "Smart Urban Management" supported by big data.The database of smart urban management system stores a large number of urban management scene images,which cover all kinds of complex street scenes and almost all types of urban management cases in the city.Therefore,how to make full use of these image data,so as to better achieve intelligent urban management,is an important research topic,with high scientific value and application value.At present,there is no public large-scale image dataset of urban management scene in the academic circle.Therefore,the primary task of this research field is to produce a large-scale and high-quality dataset.However,there is a large amount of face privacy information in urban management image.To prevent privacy leakage is a problem that must be solved in dataset disclosure.Case recognition is the most basic task in smart urban management scene understanding.It can quickly recognize the cases in the photos taken and uploaded by urban management patrol personnel or in the real-time pictures of urban surveillance cameras,so as to inform the corresponding departments to deal with them,realizing automatic and intelligent urban management.Due to the complexity of urban scenes and the small size of key objects in some cases,it is very difficult to accurately identify the types of cases.Therefore,to study and solve these problems is the key to break through the bottleneck of current algorithm recognition rate.On the basis of case recognition,it is particularly important to locate key objects,which can help achieve more efficient and accurate urban management.Traditional fully supervised object localization or detection algorithms need to spend a lot of manpower and material resources to label the object bounding box.However,the weakly supervised object localization algorithms proposed in recent years have the disadvantages of incomplete object localization and excessive background localization.The main research content of this paper is the application of generation,recognition and localization algorithms based on the deep learning in smart urban management scene understanding.The research also involves other cutting-edge technologies like the Generative Adversarial Networks(GAN)and the Weakly Supervised Learning.The main innovations are as follows:1)The Smart Urban Management image dataset UMC is built.We collect 150,000 images and proposed the first large-scale urban management image dataset.UMC covers 19 common categories of urban management cases.Moreover,we adopt image generation and processing algorithms to enhance the resolution of some low resolution images,and to deblur some blurred images,greatly improving the quality of this dataset and the performance of the recognition and location algorithms on this dataset.The experimental results show that the accuracy of recognition algorithm is improved by 0.4-0.6 percentage points,and the accuracy of location algorithm is improved by 0.7-0.8 percentage points.2)To solve the problem of face privacy protection in urban management scene,high fidelity face generation and swapping algorithms FIT-GAN and AP-GAN are proposed.They can replace the target identity in the original image or video with any specified source identity or fake identity synthesized by algorithms,thus protect the face privacy without affecting the quality and visual appearance of the image.Note that the proposed algorithms are far ahead of state-of-the-art face synthesis and swapping algorithms.3)An urban management case recognition model with bidirectional feature transfer path and multi-level classifieris is proposed.In order to solve the problems of complex urban management scene and small key objects in some cases,we propose the Multi-Level Ensemble Network(MLEN).MLEN introduces classifiers in the shallow layers(low feature levels)of the network,leveraging the multi-classifier ensemble method to get the final prediction.Meanwhile,we analyze the influence of introducing the low-level classifier on neural network,and creatively propose "Feature Transfer Path"(FTP).On the UMC dataset,MLEN has achieved 0.884 Top-1 accuracy,which is 2%higher than state-of-the-art recognition networks.In addition,on the public scene recognition dataset,MLEN-92 has achieved 0.576 Top-1 accuracy,which is 0.6%higher than the state-of-the-art network DPN-131.It reveals that MLEN can greatly improve the accuracy of case recognition with fewer parameters.4)A weakly supervised object localization algorithm with the adaptive attention augmentor is proposed.In order to solve the problem that the state-of-the-art weakly supervised localization algorithms can not locate the complete object and introduce too many backgrounds,we propose the "Adaptive Attention Augmenter"(A3),which can be easily embedded into any network classifier.It can adaptively enhance the object attention and suppress the background attention on the attention map,so as to accurately locate the object.On the ILS VRC dataset,the accuracy of A3 is more than 2 percentage points higher than the state-of-the-art Weakly Supervised Localization algorithms.On CUB-200 and Cars-196 datasets,A3 also reaches the level of state-of-the-art.On the UMC dataset,the accuracy of A3 is more than 1 percentage point higher than the state-ofthe-art Weakly Supervised Localization algorithms.In conclusion,firstly,we build a large-scale urban management image dataset called UMC,and use the super-resolution algorithm and deblur algorithm to improve its image quality.Secondly,face synthesis and swapping algorithms are proposed to effectively avoid the face privacy leaks that may exist in the dataset release without affecting the image appearance and quality.Thirdly,we propose a case recognition algorithm with the bidirectional feature transfer path and multi-level classifiers,which makes full use of multi-scale features and greatly improves the accuracy of case recognition.Finally,we propose a weakly supervised object localization algorithm with the adaptive attention augmentor,which solves the defects of the state-of-the-art methods that can not locate the integral object and introduce too many backgrounds.