Research On Quantitative Methods Of Perception Of Urban Visual Attributes

With the modernization of cities,contemporary cities are paying more and more attention to the development model of combining machine vision and life.There are also more and more researchers in urban attribute vision,because the physical and mental health of urban residents,the quality of education,and social activities all affect human visual perception of urban appearance.Scientists have begun to study urban attribute vision with the help of a combination of artificial intelligence and visual knowledge.Most of the current urban attribute visual perception analysis methods are based on the analysis of pictures by humans,and then predict the future development of the city based on the information obtained from the pictures.The limitations of the naked eye and even humans limit our ability to predict the attributes of cities.Differences between people’s subjective consciousness also affect the prediction of the city’s future.To this end,in recent years,it has been proposed to use machine vision to dynamically analyze city attributes.By setting up cameras in the urban area,the citizens’ living conditions and traffic flow are captured at all times,and then these image data are used to predict the human perception of urban attributes by means of a sorting network algorithm.All the works of this paper are to use the image’s perceptual attribute class activation map to analyze the basic information of the image,visual vocabulary information or semantic information features,and then use the obtained image information to help predict the future development of the city.With the help of this idea and corresponding theory,the basic feature information and semantic information of the image are integrated into machine learning,thereby improving our prediction accuracy of the future development of the city.A large amount of literature data confirms the feasibility and superiority of our method.One of the works of this paper is to transform the network features obtained by graphics from classification tasks to sorting tasks in the form of image comparison pairs.Doing so can avoid the limitation of human prediction.However,the features used in the previous methods are high-level features after deep convolution.In psychology,the color and texture information of the image will affect people’s perception and judgment.The previous methods have not considered the importance of the basic feature information of these images.Then we firstly proposed abrand-new sorting network model based on attention mechanism.This method uses the attention mechanism to fuse the color features and texture features that have been preprocessed from the color and texture into the feature representation,and then uses this feature to predict the perception score.We use more in-depth learning methods to quantify urban environmental perception and study the relationship between urban appearance and residents’ safety,so that we can know how to improve urban conditions.However,the current state-of-the-art methods in the world simply use convolutional neural networks to extract image feature representations from the street images of the original city without considering the cognitive factors that affect human perception of the urban environment.The second work in this paper is a continuation of the first work.The first method proposed in this paper is affected by color and texture information between different attributes,and the impact effect is large.This method requires image processing in advance.The steps are relatively cumbersome.And because the low-level features contain rich basic image information,such as color,texture,shape,and spatial relationship.High-level features include abstract information of the image,such as semantic information,low-level features and high-level features are layered together.The low-level feature space is rich in information,but lacks semantic information,and the high-level feature is the opposite.The human perception of the image is extremely rich and the perception factors are subjective.It is easy to unconsciously receive the influence of the basic information and semantic information of the image.The large amount of data collected is difficult to label one by one,which makes us unable to attribute all the attributes Preparing sufficient annotations has become a major obstacle to solving the urban attribute visual perception problem.In order to solve this problem,we have introduced a multi-level feature fusion algorithm in urban visual perception.Multi-level feature fusion is one of the target detection methods widely concerned in computer vision.We can effectively use multi-level information by combining multi-level feature training,and make full use of the basic image features and high-level semantic features that affect people’s perception,so as to obtain a more visual representation of human visual characteristics and realize the new image in the perceptual attributes On perceptual prediction.We label and train some representative data,and then use the basic feature information and object semantic information in the image to mine the association between perceptual attributes and.The comparison experiment with other ranking network methods proves the superiority of ourproposed prediction method based on multi-level feature fusion.