A Deep Learning Screening Method For Crowdsourcing Images Of Ground Objects

The development of network technology has enabled crowdsourcing geographic data to reach millions of households.As an important part of crowdsourcing geographic data,crowdsourcing images are more informative and more intuitive.Crowdsourcing imagery has the characteristics of large base,complex sources,and low correlation.Therefore,how to quickly and accurately screen out the necessary image sets in massive crowdsourcing image data is particularly important.At present,the research on the screening of crowdsourcing images often focuses only on certain aspects of the image and fails to make full use of the image information.At the same time,there are problems such as low screening accuracy and cumbersome calculations.Therefore,it is very important to study a new image screening method.In recent years,machine learning,especially deep learning,has played an increasingly important role in pattern recognition,image classification,and image processing.Convolutional neural network can mine the deep features of images,and it is highly invariant to translation,scaling,tilting and other forms of deformation.Therefore,using deep learning to screening the crowdsourcing images has become a trend.The research in this paper is aimed at the crowdsourcing images of ground objects.Since the main source of the images are collected by a smart phone or other camera device,these images are influenced by light and capture angles.The histogram of oriented gradient is invariant to the above two deformations.Based on this,this paper uses the histogram of oriented gradient as one of the expressions of the image.The histogram of oriented gradient is combined with the features extracted by the convolutional neural network to characterize the images.At the same time,this paper uses the modified cosine similarity comparison method to measure the similarity of images,overcomes the insensitivity of the cosine similarity method to numerical differences,and makes full use of the image fusion features.In this paper,a series of experiments on the crowdsourcing images of the ground objects were carried out using the above method,including the optimization experiments to determine the dimensions of histogram of oriented gradient,the experiments about the fusion feature of the histogram of oriented gradient and the feature extracted by convolutional neural network.Compare the influence of different similarity comparison algorithms on the screening results,and evaluate the screening efficiency and prevalent applicability of the method.Experiments show that the process proposed could handle the crowdsourcing image screen for multiple ground objects and achieve higher recognition rate and efficiency.